Optimizing STAR Alignment Parameters for Enhanced Novel Splice Junction Detection in Biomedical Research

Abstract

This article provides a comprehensive guide for researchers and drug development professionals on optimizing the STAR aligner for novel splice junction detection, a critical capability for identifying disease-relevant splicing variants. We cover foundational concepts of spliced alignment, detail step-by-step protocols for two-pass alignment and parameter configuration, address common troubleshooting and optimization challenges, and present rigorous validation frameworks. By integrating current methodological insights with practical optimization strategies, this resource empowers scientists to maximize sensitivity and accuracy in splicing analyses, thereby advancing transcriptomic studies in cancer, neurodegeneration, and other splicing-associated diseases.

Understanding Splice Junction Detection: Why Novel Junctions Matter in Disease Research

The Critical Role of Alternative Splicing in Disease Mechanisms and Drug Targeting

Alternative splicing (AS) is a fundamental mechanism enabling a single gene to produce multiple protein isoforms with diverse or even opposing functions, thereby greatly expanding the functional complexity of the proteome [1]. This process is critically regulated by the spliceosome, a molecular complex that acts as 'cellular scissors' to remove introns and join exons in pre-messenger RNA (pre-mRNA), with different exon combinations generating distinct mature mRNAs [1]. When alternative splicing is dysregulated, it can contribute to the pathogenesis of numerous diseases, ranging from rare genetic disorders to common complex diseases, making it a compelling target for therapeutic intervention [2] [1]. This application note explores the mechanisms linking splicing defects to disease, highlights successful drug targeting strategies, and provides detailed protocols for investigating alternative splicing in a research setting, with a specific focus on optimizing STAR aligner parameters for novel splice junction detection.

Alternative Splicing in Disease Pathogenesis

Mechanisms of Splicing Dysregulation

Dysregulation of alternative splicing can occur through multiple mechanisms, each with significant pathological consequences:

• Cis-acting mutations: Genetic variants in splice sites or regulatory elements (e.g., exonic or intronic splicing enhancers/silencers) can disrupt normal splicing patterns. For example, in Spinal Muscular Atrophy (SMA), incorrect skipping of an exon in the SMN1 gene produces a non-functional protein, leading to motor neuron loss [1].
• Trans-acting factor dysregulation: Altered expression or function of RNA-binding proteins (RBPs) and core spliceosome components can have widespread effects on splicing networks. For instance, the splicing factor Muscleblind1 depletion causes specific exon skipping in heart tissue, as visualized through Sashimi plots [3].
• Environmental stress-induced splicing changes: Cellular stress responses can directly influence splicing decisions, contributing to disease progression in conditions like inflammatory bowel disease (IBD), where immune cells under stress produce different isoforms of the BCL gene to control survival and proliferation [1].

Approximately 10-30% of disease-causing variants are estimated to affect splicing, highlighting the broad impact of splicing dysregulation on human health [1].

Disease Associations and Quantitative Evidence

Table 1: Alternative Splicing Associations in Human Diseases

Disease Category	Specific Disease	Key Gene(s)	Splicing Defect	Functional Consequence
Rare Genetic Disorder	Spinal Muscular Atrophy (SMA)	SMN1	Exon skipping	Loss of motor neurons [1]
Metabolic Disorder	Obesity, Type 2 Diabetes, Dyslipidemias	Multiple (sQTL associated)	Tissue-specific mis-splicing	Metabolic dysregulation [2]
Inflammatory Disease	Inflammatory Bowel Disease (IBD)	>200 genomic regions	Intronic variant disruption	Altered immune cell regulation [1]
Neurological Disease	Epilepsy	SCN1A (sodium channel)	Splicing variation	Altered response to antiepileptic drugs [4]

Table 2: Splicing Quantitative Trait Loci (sQTL) Impact on Complex Traits

sQTL Feature	Impact and Prevalence
Definition	Genetic loci that influence variation in alternative splicing patterns [2] [4]
Association	Strongly associated with cardiometabolic traits and disease risk [2]
Detection Method	Robust statistical models like GLiMMPS using RNA-seq data [4]
Validation Rate	100% validation rate for 26 randomly selected sQTLs via RT-PCR [4]

Therapeutic Targeting of Alternative Splicing

Approved Splicing-Targeted Therapies

The most prominent success in splicing-directed therapeutics is Nusinersen (Spinraza), an FDA-approved antisense oligonucleotide for Spinal Muscular Atrophy that corrects the aberrant skipping of exon 7 in the SMN1 gene, transforming a fatal condition into a manageable disorder [1]. This therapy exemplifies the principle of using antisense oligonucleotides to modulate splicing outcomes by binding to specific pre-mRNA sequences and blocking or promoting the inclusion of specific exons.

Emerging Therapeutic Approaches

• RNA-based medicines: Following the success of mRNA vaccines, RNA-targeted therapeutics are emerging as a promising modality for correcting splicing errors, with several now approved for clinical use [1].
• Small molecule splicing modulators: Compounds that target specific components of the splicing machinery or regulatory complexes offer potential for oral administration and broader application.
• Population-scale splicing maps: Initiatives like the IsoIBD project and Project JAGUAR are building detailed maps of splicing variation in disease-relevant tissues across diverse populations, providing foundations for targeted treatment development [1].

Therapeutic Development Pipeline

Experimental Protocols for Splicing Analysis

RNA Sequencing and Splice Junction Detection

The accurate detection of splice junctions from RNA-seq data presents significant computational challenges, particularly in distinguishing between biologically relevant isoforms and technical artifacts [5]. The STAR (Spliced Transcripts Alignment to a Reference) aligner uses a novel algorithm based on sequential maximum mappable seed search in uncompressed suffix arrays followed by seed clustering and stitching, enabling precise mapping of reads across splice junctions without prior knowledge of junction locations [6].

Table 3: STAR Aligner Parameter Optimization for Splice Junction Detection

Parameter	Standard Setting	Optimized for Novel Junctions	Impact on Detection
Alignment Mode	1-pass mapping	2-pass mapping with filtered junctions	Increases novel junction discovery [7]
Junction Filtering	None	Remove low-coverage (column 7 < 5) and non-canonical junctions	Improves reproducibility [7]
Genome Alignment	Basic genome index	Include comprehensive annotation	Reduces false positives [6]
Read Mapping	Default parameters	Increase alignments per read for multimapping	Enhances sensitivity [6]

Protocol: Two-Pass STAR Alignment with Junction Filtering

Materials Required:

• RNA-seq data in FASTQ format
• Reference genome and annotation (GTF/GFF)
• STAR aligner software
• Computing resources (12-core server recommended for efficiency [6])

Procedure:

• Generate genome index: Create a STAR genome index using the reference genome and annotation file.
• First-pass alignment: Run STAR in first-pass mode to identify novel splice junctions for each sample.
• Junction filtering: Concatenate SJ.out.tab files from all samples and filter junctions with low read support (<5 reads), non-canonical splice sites, and mitochondrial genes.
• Second-pass alignment: Re-run STAR using the filtered junction file from step 3 to improve alignment accuracy and novel junction detection.
• Output analysis: Process the resulting BAM files for downstream splicing quantification using tools like MAJIQ or MISO.

Note: While 2-pass mapping increases detection of splicing changes, it may reduce uniquely mapped reads by 1-2% and increase computational time. For most applications, 1-pass mapping provides more reproducible results, while 2-pass is preferable for hypothesis-generating studies seeking maximal junction discovery [7].

Splicing Quantification and Statistical Analysis

The GLiMMPS (Generalized Linear Mixed Model Prediction of sQTL) method provides a robust statistical framework for detecting splicing quantitative trait loci (sQTLs) from RNA-seq data, explicitly accounting for individual variation in sequencing coverage and overdispersion prevalent in RNA-seq data [4]. Unlike simple linear models, GLiMMPS models the estimation uncertainty of exon inclusion levels (PSI, Percent Spliced In) by using reads from both inclusion and skipping isoforms, significantly improving detection reliability with a demonstrated 100% validation rate for identified sQTLs [4].

RNA-seq Splicing Analysis Workflow

Visualization and Interpretation

Sashimi Plots for Quantitative Splicing Visualization

Sashimi plots provide a quantitative multi-sample visualization of RNA sequencing read alignments, enabling direct comparison of splicing patterns across different conditions or genotypes [3]. These plots display genomic reads as density plots (in RPKM units) and splice junction reads as arcs whose width is proportional to the number of junction reads spanning connected exons, with raw junction read counts annotated on each arc [3].

Protocol: Generating Sashimi Plots with IGV and Command-Line Tools

• Input Requirements: Alignments in BAM format and gene model annotations in GFF3 format.
• IGV-Sashimi Integration: Use the Integrated Genome Browser (IGV) for dynamic, exploratory analysis of genomic regions with on-the-fly Sashimi plot generation.
• Static Sashimi Plots: Use the command-line Python implementation (packaged with MISO) for publication-quality figures with customizable colors, scales, fonts, and sizes.
• Quantitative Integration: Optionally decorate plots with PSI (Ψ) values from MISO estimation to provide quantitative measures of alternative exon inclusion levels.

Long-Read Sequencing for Comprehensive Isoform Characterization

While short-read sequencing (75-150 base pairs) has been the standard for RNA studies, long-read technologies from Pacific Biosciences and Nanopore now enable sequencing of full-length RNA molecules spanning thousands of base pairs, providing unambiguous information about alternative splicing structure without assembly [1]. This is particularly valuable for clinical applications where accurately determining the complete structure of splicing variants is essential for understanding pathogenic mechanisms and developing targeted interventions.

The Scientist's Toolkit: Research Reagent Solutions

Table 4: Essential Research Reagents and Computational Tools for Splicing Analysis

Reagent/Tool	Category	Function and Application	Key Features
STAR Aligner [6]	Computational Tool	Spliced alignment of RNA-seq reads to reference genome	Ultra-fast, detects canonical and non-canonical junctions
GLiMMPS [4]	Statistical Model	Detection of splicing quantitative trait loci (sQTLs)	Accounts for read depth variation and overdispersion
Sashimi Plots [3]	Visualization	Quantitative visualization of splicing across samples	Junction read arcs and read density profiles
Pacific Biosciences [1]	Sequencing Technology	Long-read sequencing for full-length isoform detection	Resolves complex splicing patterns without assembly
MISO [3]	Computational Tool	Quantification of alternative splicing from RNA-seq data	Bayesian estimation of isoform abundance (PSI values)
MAJIQ [7]	Computational Tool	Detection and quantification of splicing changes	Models local splicing variations (LSVs) and confidence
Rosthornin B	Rosthornin B, MF:C24H34O7, MW:434.5 g/mol	Chemical Reagent	Bench Chemicals
6-O-Cinnamoylcatalpol	6-O-Cinnamoylcatalpol	High-purity 6-O-Cinnamoylcatalpol for research applications. Explore its potential anti-inflammatory and anti-tumor properties. This product is For Research Use Only (RUO). Not for human or veterinary use.	Bench Chemicals

Alternative splicing represents a critical layer of genetic regulation with profound implications for understanding disease mechanisms and developing targeted therapies. The integration of advanced computational methods like STAR alignment with robust statistical approaches such as GLiMMPS enables comprehensive detection and quantification of splicing variations associated with disease. As long-read sequencing technologies mature and population-scale splicing maps expand, researchers and drug developers are increasingly equipped to identify splicing-specific therapeutic targets and develop innovative RNA-directed medicines that can correct pathogenic splicing defects across a wide spectrum of human diseases.

The accurate detection of novel splice junctions from RNA sequencing (RNA-seq) data is crucial for advancing our understanding of transcriptome complexity, particularly in disease research and drug development. However, standard alignment methodologies exhibit inherent biases that favor known junctions, impeding the discovery of unannotated splicing events. Within the context of optimizing STAR (Spliced Transcripts Alignment to a Reference) alignment parameters, this application note delineates the primary challenges and provides detailed protocols to overcome these technical limitations, enabling more sensitive and accurate novel junction detection.

The Core Challenge: Alignment Biases Against Novel Junctions

Conventional RNA-seq aligners, including STAR, typically use annotated gene references to facilitate the alignment process. This practice creates a systematic bias where alignment algorithms require substantially more evidence to align reads across novel splice junctions compared to known, annotated junctions [8]. The preference for known junctions is often implemented through varied alignment scores or multi-stage alignment processes, which inadvertently reduces sensitivity for discovering novel biological events [8]. This bias directly impacts the detection of clinically significant splice-altering somatic variants, which are crucial for predicting treatment risk and making therapeutic decisions in hematologic malignancies [9].

Solution: Two-Pass Alignment with STAR

The two-pass alignment method effectively separates the discovery of splice junctions from their quantification. In the first pass, junctions are identified with high stringency. These discovered junctions are then used as a custom "annotation" in the second alignment pass, which is performed with lower stringency to permit higher sensitivity for aligning reads to these now-known novel junctions [8].

Detailed Experimental Protocol

1. First-Pass Alignment for Junction Discovery

Note: The --alignSJoverhangMin 8 parameter sets a high stringency for novel junction discovery in the first pass [8].

2. Genome Re-indexing with Discovered Junctions

3. Second-Pass Alignment for Sensitive Quantification

Note: The --alignSJDBoverhangMin 3 parameter in the second pass allows reads to span splice junctions with fewer nucleotides, significantly increasing sensitivity [8].

Performance Characteristics

The following table summarizes the quantitative improvements observed with two-pass alignment across various RNA-seq datasets:

Table 1: Performance Metrics of Two-Pass Alignment Across Diverse RNA-seq Datasets

Sample Type	Read Length	Junctions Improved	Median Read Depth Ratio	Expected Read Depth Ratio
Lung Adenocarcinoma Tissue	48 nt	99%	1.68×	1.75×
Lung Normal Tissue	48 nt	98%	1.71×	1.75×
Reference RNA (UHRR)	75 nt	94-97%	1.25-1.26×	1.35×
Lung Cancer Cell Lines	101 nt	97%	1.19-1.21×	1.23×
Arabidopsis Samples	101 nt	95-97%	1.12×	1.12×

Data adapted from systematic evaluation of two-pass alignment performance [8].

Critical STAR Parameters for Novel Junction Detection

Parameter Optimization Strategy

Optimizing specific STAR parameters is essential for balancing sensitivity and precision in novel junction detection. The following parameters have demonstrated significant impact on performance:

Table 2: Key STAR Alignment Parameters for Optimizing Novel Junction Detection

Parameter	Recommended Setting	Impact on Performance	Biological Rationale
--alignSJoverhangMin	8 (1st pass), 3 (2nd pass)	Higher values increase precision; lower values increase sensitivity	Prevents false positives while enabling detection of junctions with short overhangs
--alignIntronMin	20	Reduces false positives from small indels	Matches minimum known intron size in eukaryotes
--alignIntronMax	1000000	Allows discovery of long-range splicing	Accommodates the longest known human introns
--outFilterType	BySJout	Reduces false junctions in output	Filters out alignments with spurious splice junctions
--scoreGenomicLengthLog2scale	0	Eliminates intron length bias	Prevents penalization of longer introns during alignment

Experimental Workflow for Parameter Optimization

The following diagram illustrates the logical workflow for optimizing STAR parameters to address alignment biases:

Validation and Quality Control Measures

Addressing Potential Alignment Errors

While two-pass alignment significantly improves sensitivity, it may introduce alignment errors. These potential errors are readily identifiable through simple classification methods and can be filtered using the following criteria:

1. Multi-mapping Read Filtering

• Discard junctions supported predominantly by reads that map to multiple genomic locations
• Require at least 2-3 uniquely mapping reads per novel junction

2. Sequence Motif Validation

• Verify presence of canonical splice motifs (GT-AG, GC-AG, AT-AC)
• Flag non-canonical junctions for additional scrutiny

3. Experimental Validation Protocol For high-priority novel junctions, experimental validation remains the gold standard:

This approach has demonstrated success rates of 80-90% in validating novel intergenic splice junctions [6].

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Research Reagents and Computational Tools for Splice Junction Studies

Resource	Type	Primary Function	Application Context
STAR Aligner	Software	Spliced alignment of RNA-seq reads	Ultrafast mapping of spliced sequences of any length [6] [10]
GENCODE Annotation	Database	Comprehensive gene and transcript annotation	Reference for known splice junctions and gene models [8]
Two-pass Alignment	Protocol	Enhanced novel junction quantification	Increases sensitivity for unannotated splicing events [8]
ENCODE Guidelines	Framework	Standardized parameter settings	Ensures reproducibility across experiments [8]
PolyA Site Databases	Resource	Annotated polyadenylation sites	Context for 3' UTR splicing and alternative polyadenylation [11]
16-Epivoacarpine	16-Epivoacarpine, MF:C21H24N2O4, MW:368.4 g/mol	Chemical Reagent	Bench Chemicals
Spiramilactone B	Spiramilactone B|Supplier	Spiramilactone B (CAS 180961-65-3) is a diterpenoid for research. This product is For Research Use Only and is not intended for diagnostic or personal use.	Bench Chemicals

The implementation of optimized two-pass alignment with carefully tuned STAR parameters effectively mitigates the inherent biases against novel splice junction detection. The detailed protocols provided herein enable researchers to achieve as much as 1.7-fold improvement in read coverage over novel junctions, significantly enhancing the discovery of biologically and clinically relevant splicing events. This approach is particularly valuable in oncology and drug development contexts where splice-altering variants represent important therapeutic targets and biomarkers.

The fundamental challenge of RNA-seq read alignment, compared to genomic DNA alignment, stems from the non-contiguous nature of mature messenger RNA (mRNA) transcripts. In eukaryotic cells, pre-mRNA undergoes splicing where introns are removed and exons are joined together to form the final mRNA molecule. Consequently, reads derived from RNA-seq experiments may span these splice junctions, meaning one portion of the read aligns to one exon while the remaining portion aligns to a non-adjacent exon separated by a potentially large intron in the reference genome. A splice-aware aligner like STAR is specifically designed to handle this biological reality by not attempting to align RNA-seq reads contiguously across introns and instead identifying possible downstream exons to align to, effectively ignoring introns altogether [12].

In contrast, traditional DNA-DNA aligners are considered "splice-unaware." When encountering a read that spans a splice junction, such an aligner would need to introduce a very long gap in the alignment to bridge the intron. This is computationally undesirable and often leads to false mappings, as the aligner might instead find an incorrect, contiguous genomic sequence that partially matches the read [12]. Therefore, while it is possible to align RNA-seq reads to a reference transcriptome using a splice-unaware aligner, aligning to the full genome—which enables the discovery of unannotated genes and novel splice junctions—absolutely requires a splice-aware aligner [12].

The STAR Algorithm: Core Mechanics

STAR (Spliced Transcripts Alignment to a Reference) employs a unique two-step strategy that sets it apart from traditional aligners and underpins its high speed and accuracy [6] [13].

Seed Searching with Maximal Mappable Prefixes (MMPs)

For each read, STAR performs a sequential search to find the longest subsequence that exactly matches one or more locations on the reference genome [6] [13]. These longest matches are called Maximal Mappable Prefixes (MMPs). The algorithm starts from the beginning of the read, finds the first MMP (designated seed1), then repeats the search for the unmapped portion of the read to find the next MMP (seed2). This process continues until the entire read is processed [13]. This sequential search of only the unmapped portions is a key factor in STAR's efficiency. The MMP search is implemented using uncompressed suffix arrays (SAs), which allow for rapid searching against large reference genomes with logarithmic scaling of search time relative to genome size [6]. If a read contains mismatches or indels that prevent an exact match, the previously identified MMPs are extended to accommodate the differences. If extension fails, poor quality or adapter sequences are soft-clipped [13].

Clustering, Stitching, and Scoring

In the second phase, STAR builds complete alignments by stitching the individual seeds together. The seeds are first clustered based on their proximity to a set of "anchor" seeds, which are seeds that map uniquely to the genome. A dynamic programming algorithm then stitches the seeds within a cluster, allowing for mismatches and a single insertion or deletion (gap). The final alignment is selected based on a scoring model that accounts for mismatches, indels, and gaps [6] [13]. For paired-end reads, STAR clusters and stitches seeds from both mates concurrently, treating the read pair as a single sequence. This approach increases sensitivity, as only one correct anchor from one mate is often sufficient to accurately align the entire read pair [6].

Table 1: Core Algorithmic Differences Between Spliced and Genomic Alignment

Feature	Splice-Aware Aligner (STAR)	Standard Genomic Aligner
Handling of Spliced Reads	Aligns segments to separate exons; identifies splice junctions	Attempts contiguous alignment; often fails or misaligns
Reference Requirement	Genome (preferred) or transcriptome	Genome or transcriptome
Key Innovation	Maximal Mappable Prefix (MMP) search & seed stitching	Full-length or seed-and-extend alignment
Gap Handling	Interprets large gaps as introns; uses splice junction penalties	Interprets large gaps as potential indels
Output Capabilities	Identifies canonical & novel splice junctions, chimeric transcripts	Identifies genomic variants (SNPs, indels)

Two-Pass Alignment for Novel Junction Discovery

A critical protocol for enhancing the detection of novel splice junctions is the two-pass alignment method [8] [14]. This strategy is particularly valuable for research aimed at comprehensive splice junction discovery, such as in the context of the user's broader thesis on STAR parameters.

Protocol Rationale and Workflow

In standard single-pass alignment, the aligner uses existing gene annotations to guide the mapping of reads across known splice junctions. While this reduces noise, it also introduces a bias against the alignment of reads that span novel, unannotated junctions, as these require more evidence than known junctions [8]. The two-pass method addresses this by separating the processes of junction discovery and sensitive read quantification [8].

The workflow consists of two sequential alignment steps:

• First Pass: Alignment is performed with high stringency using only the existing genome and annotations. The goal is to generate a high-confidence set of novel splice junctions from the data, which are recorded in the SJ.out.tab output file.
• Second Pass: The novel junctions discovered in the first pass are provided to STAR as an additional set of "annotations" via the --sjdbFile option. The data is then realigned. In this pass, reads can be mapped across these novel junctions with the same lower stringency typically reserved for known junctions, thereby increasing sensitivity and the accuracy of quantification for these novel sites [8] [14].

Performance and Considerations

Two-pass alignment has been demonstrated to significantly improve the quantification of novel splice junctions. Studies show it can improve the median read depth over novel junctions by as much as 1.7-fold, with 94-99% of simulated novel junctions being more accurately quantified compared to single-pass alignment [8].

However, this increased sensitivity comes with trade-offs that must be considered for a research project. It can lead to a 1-2% decrease in the percentage of uniquely mapped reads and a substantial increase in run time, especially with large datasets that generate many novel junction annotations [7]. Furthermore, while two-pass mapping detects more splicing changes, the additional splicing events identified may be less reproducible than those found with single-pass mapping [7]. To mitigate some of these issues, it is recommended to filter the junctions from the first pass before the second pass, removing junctions with low read support (e.g., < 5 reads), non-canonical motifs, and those on the mitochondrial chromosome [7].

Table 2: Two-Pass vs. One-Pass Alignment for Splice Junction Detection

Characteristic	One-Pass Alignment	Two-Pass Alignment
Primary Goal	Efficient mapping with reference annotation	Maximized discovery of novel splice junctions
Sensitivity to Novel Junctions	Lower (biased towards annotated junctions)	Higher (treats novel junctions as known in second pass)
Quantification Accuracy	Good for annotated junctions	Improved for novel junctions (up to 1.7x read depth)
Computational Load	Faster, less resource-intensive	~3-5 minutes more per sample; requires more storage
Uniquely Mapped Reads	Higher percentage	0.4% - 2% lower
Best Use Case	Standard gene expression quantification, well-annotated organisms	Exploratory research, novel isoform detection, less-studied genomes

Experimental Protocol for STAR Alignment

This protocol provides a detailed methodology for generating genome indices and performing read alignment with STAR, forming the basis for reproducible RNA-seq analysis [13] [14].

Hardware

• Computer with Unix, Linux, or Mac OS X.
• Substantial RAM: at least 10 x genome size (e.g., ~30 GB for human). 32 GB is recommended for human genomes [14].
• Sufficient disk space (>100 GB) for output files and temporary storage.
• Multiple CPU cores for parallel processing. The number of threads (--runThreadN) is typically set to the number of physical cores.

Software

• STAR software, available from https://github.com/alexdobin/STAR [14].

Input Files

• Reference genome sequence in FASTA format.
• Gene annotation in GTF format (e.g., from Ensembl or GENCODE).
• RNA-seq reads in FASTQ format (uncompressed or gzipped).

Step-by-Step Procedures

Alternate Protocol 1: Generating Genome Indices Genome indices must be created once for each genome/annotation combination.

• Create Directory and Navigate:

• Run Genome Generation: Execute the STAR command in genomeGenerate mode. The --sjdbOverhang should be set to the maximum read length minus 1 [13].

Basic Protocol: Mapping RNA-seq Reads This is the core mapping procedure, which can be performed as a single-pass or as the first pass of a two-pass strategy.

• Create and Navigate to Run Directory:

• Execute Alignment Job: The command below is for paired-end reads. For single-end, specify only one file after --readFilesIn. The --outSAMtype option specifies a coordinate-sorted BAM file, which is standard for downstream analysis.

Alternate Protocol 2: Two-Pass Mapping To perform two-pass mapping for novel junction discovery [14]:

• First Pass: Run the Basic Protocol alignment. This generates a file sample1_SJ.out.tab containing discovered junctions.
• Second Pass: Run the Basic Protocol again, but add the junctions from the first pass as additional input.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Reagents for STAR RNA-seq Analysis

Item	Function/Description	Example Source/Format
Reference Genome	Linear genomic sequence for read alignment; the foundational mapping coordinate system.	FASTA file (e.g., GRCh38 from GENCODE/Ensembl)
Gene Annotation	Provides known transcript models and splice junctions to guide the initial alignment.	GTF/GFF3 file (e.g., from GENCODE, Ensembl, or RefSeq)
RNA-seq Reads	The experimental data; short sequence fragments derived from fragmented mRNA.	FASTQ files (single- or paired-end, gzipped or uncompressed)
Splice Junction Database (e.g., SJ.out.tab)	A list of high-confidence, sample-derived junctions used as custom annotation for the second pass of alignment.	Tab-delimited file generated by a first-pass STAR run
High-Performance Computing (HPC) Environment	Essential for handling the significant memory (~30 GB for human) and multi-core processing requirements of STAR.	Unix/Linux server or cluster with >=32 GB RAM and multiple cores
Pseudolaric acid D	Pseudolaric acid D, MF:C20H30O3, MW:318.4 g/mol	Chemical Reagent
Neolinine	Neolinine, MF:C23H37NO6, MW:423.5 g/mol	Chemical Reagent

Visualizing the STAR Alignment Algorithm

The following diagram illustrates the core two-step process of the STAR aligner, from seed search to final stitched alignment.

STAR's Two-Phase Spliced Alignment Process

Implementing Two-Pass STAR Alignment: Step-by-Step Protocols and Parameter Configuration

The two-pass alignment workflow is a sophisticated bioinformatic method designed to enhance the detection and quantification of novel splice junctions from RNA sequencing (RNA-seq) data. In standard, single-pass alignment, computational tools align reads to a reference genome using existing gene annotations, which inherently biases the process towards known splice junctions and requires substantially more evidence to align reads across novel, unannotated junctions [8]. This presents a significant challenge for research focused on discovering novel splicing events, such as in disease modeling or non-model organism studies.

The two-pass approach elegantly overcomes this limitation by separating the processes of splice junction discovery and read quantification [8]. In the first pass, alignment is performed with high stringency to identify a comprehensive set of splice junctions from the data itself. These newly discovered junctions are then used as a custom annotation to guide a second, more sensitive alignment pass. This method effectively levels the playing field, allowing novel junctions to be penalized similarly to known ones, thereby increasing sensitivity without compromising specificity. Originally developed for short-read technologies, this approach has been successfully adapted for long-read sequencing platforms like Oxford Nanopore Technologies (ONT) and Pacific Biosciences (PacBio), where higher error rates can further complicate accurate splice junction detection [15].

Theoretical Foundation and Performance

Rationale and Mechanism

The core rationale behind two-pass alignment is to reduce the alignment penalty bias against novel splice junctions. In a typical single-pass alignment, algorithms like STAR (Spliced Transcripts Alignment to a Reference) give preference to reads that align across known, annotated splice junctions [8]. This means a read spanning a novel junction must meet a higher burden of proof—often requiring longer and more perfect sequence matches—to be aligned correctly. The two-pass process mitigates this by treating junctions discovered in the first pass as "known" in the second pass.

The performance improvement primarily stems from an increased ability to align reads that have shorter sequence lengths spanning the splice junctions [8]. In practical terms, two-pass alignment has been demonstrated to permit alignment of reads with shorter overhangs at novel splice junctions, which would otherwise be rejected in a conventional single-pass approach. For long-read sequencing data, tools like 2passtools further refine this concept by applying machine-learning-based filters to the junctions discovered in the first pass, removing spurious alignments before the second pass to increase the overall accuracy of intron detection [15].

Quantitative Performance Gains

Empirical studies across diverse RNA-seq datasets have consistently demonstrated the substantial benefits of the two-pass alignment method. The following table summarizes key performance metrics from an evaluation of twelve public RNA-seq samples, showcasing the widespread applicability of the technique.

Table 1: Performance of Two-Pass Alignment Across RNA-Seq Datasets [8]

Sample	Description	Read Pairs (millions)	Splice Junctions Improved	Median Read Depth Ratio
TCGA-50‐5933_T	Lung Adenocarcinoma Tissue	48	99%	1.68×
TCGA-50‐5933_N	Lung Normal Tissue	52	98%	1.71×
UHRR_rep1	Reference RNA	83	94%	1.25×
UHRR_rep2	Reference RNA	85	97%	1.26×
LCS22T	Lung Adenocarcinoma Tissue	52	98%	1.20×
LCS22N	Lung Normal Tissue	35	96%	1.18×
A549	Lung Cancer Cell Line	92	97%	1.21×
NCI-H1437	Lung Cancer Cell Line	76	97%	1.19×
AT_flowerbuds	Arabidopsis Flower Buds	192	97%	1.12×
AT_leaves	Arabidopsis Leaves	202	95%	1.12×

The data reveals that two-pass alignment improved the quantification accuracy for 94% to 99% of simulated novel splice junctions across all tested samples, including human and Arabidopsis thaliana data [8]. The median read depth over these novel junctions increased by a factor of 1.12× to 1.71×, providing significantly greater power for downstream analysis and validation.

For long-read data, the benefits are similarly pronounced. In a study on Arabidopsis nanopore Direct RNA Sequencing (DRS) data, using reference splice junctions to guide minimap2 alignment (a form of two-pass principle) resulted in 92.1% of simulated reads aligning to the correct transcript isoform for a challenging locus (FLM gene), compared to only 19.3% with standard alignment and 40.3% with post-alignment correction tools like FLAIR [15].

Experimental Protocols

Protocol 1: Two-Pass Alignment with STAR for Short Reads

This protocol is optimized for Illumina short-read RNA-seq data and utilizes the STAR aligner, which is designed for high sensitivity and speed [16] [17].

A. Software and Data Preparation

• STAR Aligner: Ensure STAR (version 2.4.0h1 or later) is installed and available in your PATH [8] [17].
• Reference Genome: Obtain the reference genome sequence (e.g., GRCh38 for human, TAIR10 for Arabidopsis) in FASTA format [8] [17].
• Annotation File (Optional): For an enhanced first pass, a high-quality annotation file (e.g., GENCODE-Basic for human) in GTF format can be used [8] [17].
• RNA-seq Reads: Provide the sequencing reads in FASTQ format. For paired-end reads, you will have two files per sample [17].

B. Generating the Genome Index

The genome index must be generated once for a given reference genome and annotation combination.

• Create a directory for the genome indices: mkdir genomeDir
• Run the STAR genome generation command:
  Note: The --sjdbOverhang value should be set to the read length minus 1 [17].

C. First-Pass Alignment and Junction Discovery

The goal of this step is to map the reads and extract a comprehensive set of splice junctions from your data.

• Run the first alignment pass:

• Upon completion, STAR will generate a splice junction file named sample1_pass1_SJ.out.tab. This file contains the coordinates and structural information for all splice junctions detected in the first pass.

D. Second-Pass Alignment with Discovered Junctions

In this critical step, the junctions discovered in the first pass are used to create a refined genome index for the final, more sensitive alignment.

• Re-generate the genome index, this time including the novel junctions from the first pass:
  Note: Multiple junction files from several samples can be combined at this stage using a space-separated list with the --sjdbFile parameter for a unified project-level analysis.
• Run the second and final alignment pass using the new index:
  The final, high-quality alignments will be in sample1_pass2_Aligned.sortedByCoord.out.bam.

Protocol 2: Two-Pass Alignment for Long Reads using 2passtools

This protocol is designed for long-read technologies (ONT or PacBio) and uses 2passtools, which incorporates machine learning to filter splice junctions, addressing the higher error rates of these platforms [15].

A. Software and Data Preparation

• 2passtools: Install the software suite from https://github.com/bartongroup/2passtools [15].
• Minimap2: Ensure this aligner is available, as it is often used within the workflow [15].
• Reference Genome: Genome sequence in FASTA format.
• Long-read RNA-seq Data: Reads in FASTQ format.

B. Workflow Execution

• Perform First-Pass Alignment: Map the reads to the genome using a long-read aware aligner like minimap2.
• Filter Splice Junctions: Use 2passtools to process the first-pass alignments. It applies a combination of alignment metrics and a logistic regression (LR) model to filter out spurious splice junctions, retaining only high-confidence junctions [15]. This step leverages biological sequence signatures (e.g., GU-AG intron motifs) to distinguish genuine junctions.
• Perform Second-Pass Alignment: Use the filtered set of high-confidence junctions to guide the realignment of the original reads. This yields a final BAM file with significantly improved accuracy in intron identification and transcript isoform structure [15].

Workflow Visualization

The following diagram provides a consolidated overview of the two-pass alignment logic, applicable to both short-read and long-read variations of the workflow.

Two-Pass Alignment Workflow Logic

Table 2: Key Resources for Two-Pass Alignment Experiments

Item	Function / Relevance	Example / Specification
STAR Aligner	A fast, sensitive RNA-seq read mapper specifically designed for splice-aware alignment. It is a standard tool for implementing two-pass workflows with short reads [8] [16] [17].	https://github.com/alexdobin/STAR; Version 2.4.0h1 or later [8].
2passtools	A software package designed for two-pass alignment of long-read RNA-seq data, using machine learning to filter spurious splice junctions [15].	https://github.com/bartongroup/2passtools
Reference Genome	The canonical DNA sequence of the organism under study, serving as the reference for read alignment.	GRCh38 (human), TAIR10 (Arabidopsis) [8].
Annotation File (GTF)	A file containing known gene models and splice junctions. Used for initial indexing or as a benchmark [8] [17].	GENCODE-Basic (human), Ensembl, or RefSeq.
Splice Junction File	A file output by aligners like STAR listing discovered splice junctions. This is the key artifact passed from the first to the second alignment step [8].	File format: SJ.out.tab
High-Performance Computing (HPC) Cluster	Alignment, especially with two passes and large genomes, is computationally intensive and requires significant memory and processing power [17].	Recommended: 32+ GB RAM, multiple CPU cores.

Table of Contents

• Introduction
• Key Parameters and Their Mechanisms
• Experimental Protocols
• Performance and Validation Data
• The Scientist's Toolkit

The discovery of novel splice junctions from RNA-seq data is a fundamental step in understanding transcriptome diversity, with significant implications for basic research and drug discovery. The STAR (Spliced Transcripts Alignment to a Reference) aligner is a widely used tool for this task, but its performance is highly dependent on the configuration of critical parameters that govern sensitivity and precision. Misconfiguration can lead to either a failure to detect genuine biological events or an overload of false positives, compromising downstream analyses. This application note details the core parameters—--sjdbOverhang, --alignSJoverhangMin, and key filtering strategies—within the context of a broader research thesis on optimizing novel junction detection. We provide structured quantitative data, step-by-step protocols from cited studies, and visual workflows to equip researchers with the knowledge to reliably uncover novel splicing events.

Key Parameters and Their Mechanisms

Proper configuration of STAR parameters is essential to balance the sensitivity of novel junction discovery with the accuracy of alignment. The parameters --sjdbOverhang and --alignSJoverhangMin are particularly crucial, operating at different stages of the alignment process.

--sjdbOverhang: Genome Indexing for Junction Sensitivity

• Function and Workflow Stage: This parameter is used exclusively during the genome indexing step (--runMode genomeGenerate). It determines how many exonic bases from the donor and acceptor sites are used to construct a database of splice junction sequences that are added to the genome reference [18] [19].
• Mechanism: For each annotated junction, STAR creates a new sequence in the genomic reference by concatenating N bases from the end of the upstream exon with N bases from the start of the downstream exon, where N is the value specified by --sjdbOverhang [19]. This expanded reference allows reads to map across known (and later, discovered) junctions more sensitively.
• Setting the Value:
  • Ideal Scenario: The manual states the ideal value is mate_length - 1 [18] [19]. For example, for 100 bp paired-end reads, the ideal value is 99.
  • Practical Guidance: For read lengths below 50 bp, adhering to read_length - 1 is strongly recommended [19]. For longer reads (e.g., 75 bp, 100 bp, 150 bp), using a generic value of 100 is safe, efficient, and recommended by the STAR developer, as it works effectively without needing to re-index for every slight variation in read length [19]. Using a value that is too long is safer than one that is too short [19].

--alignSJoverhangMinvs.--alignSJDBoverhangMin: Alignment-Time Filtering

• Function and Workflow Stage: These parameters are used during the read alignment step to define the minimum number of bases a read must map to each side of a splice junction for the alignment to be accepted.
• Mechanism and Differentiation:
  • --alignSJDBoverhangMin: This parameter sets the minimum overhang for junctions that are already annotated in the supplied database (sjdb) or discovered in a first pass and used in the second [18]. The default value is 3.
  • --alignSJoverhangMin: This parameter sets the minimum overhang for novel (unannotated) junctions. The default value is 5, reflecting the higher stringency required for non-canonical events [8].
• Impact and Tuning: Lowering these values increases sensitivity, allowing the detection of junctions supported by reads with very short alignments, but at the risk of increasing false positives. Raising them enhances specificity. Research has shown that two-pass alignment improves the quantification of novel junctions by specifically permitting alignments with shorter overhangs that would otherwise be filtered out [8]. For microexons (very small internal exons), --alignSJDBoverhangMin does not apply if both flanking junctions are annotated [20].

The following diagram illustrates how these key parameters function at different stages of the STAR workflow for junction discovery.

Experimental Protocols

Here, we detail two foundational experimental protocols cited in the literature for optimizing novel junction discovery with STAR.

Protocol 1: Two-Pass Alignment for Novel Junction Discovery and Quantification

This protocol, adapted from [8] [14], separates the discovery and quantification of splice junctions to increase sensitivity for novel events.

• Principle: Splice junctions are discovered in a first alignment pass with high stringency. These newly discovered junctions are then used as an augmented annotation in a second alignment pass, allowing lower stringency alignment and higher sensitivity for quantification [8].
• Steps:
  • First Pass Alignment:
    • Run STAR in the default single-pass mode using a standard set of annotations (e.g., GENCODE or Ensembl GTF).
    • Critical output: The SJ.out.tab file from this pass contains a list of detected junctions, both annotated and novel.
  • Junction Filtering (Recommended):
    • To prevent an accumulation of false positives in the second pass, filter the SJ.out.tab file. Common filters include [7]:
      • Removing junctions with low read support (e.g., column 7, read_count < 5).
      • Keeping only canonical junctions (column 5, intron_motif = 1) if the study focus is on major splicing events.
      • Removing junctions on mitochondrial DNA (chrM).
  • Second Pass Alignment:
    • Re-run STAR on the same reads, but this time include the filtered SJ.out.tab file from the first pass using the --sjdbFileChrStartEnd parameter. This directs STAR to treat these junctions as "known" in the second pass.
    • Alternatively, the --twopassMode Basic option can be used to perform these steps automatically, though it offers less control over junction filtering [21] [14].
• Key Parameters from [8]:
  • --alignSJoverhangMin 8
  • --alignSJDBoverhangMin 3
  • --seedSearchStartLmax 30 (for increased sensitivity with shorter reads)

Protocol 2: Post-Alignment Junction Filtration with Portcullis

This protocol, based on [22], addresses the high false-positive rate of junction callers by using a dedicated filtration tool.

• Principle: Portcullis analyzes the set of mapped split reads supporting each SJ from a STAR BAM file, combining multiple metrics (e.g., read stacking, genome sequence at splice sites) to distinguish genuine from invalid junctions [22].
• Steps:
  • Initial Alignment: Run STAR in a standard single-pass mode to generate a BAM file.
  • Run Portcullis: Execute Portcullis with the BAM file and the reference genome as input.
    • portcullis full -t <threads> <reference_genome> <input.bam> <output_prefix>
  • Output: Portcullis produces a high-confidence set of junctions in various formats (BED, GTB) suitable for downstream analyses or for use as a filtered junction database in a two-pass STAR workflow.
• Performance: Independent benchmarking shows that Portcullis significantly improves precision over raw STAR output, maintaining F1 scores above 97% across different species and read lengths, even as sequencing depth increases [22].

Performance and Validation Data

The following tables summarize key quantitative findings from studies that evaluated the impact of different STAR parameters and workflows on junction discovery.

Table 1: Performance of Two-Pass Alignment on Novel Junction Quantification [8] This study treated known junctions as unannotated to simulate novel junction discovery, comparing one-pass versus two-pass alignment across diverse RNA-seq samples.

Sample Type	Read Length	Junctions Improved	Median Read Depth Ratio (2-pass / 1-pass)
Lung Adenocarcinoma	48 nt	99%	1.68×
Reference RNA (UHRR)	75 nt	94%	1.25×
Lung Cell Lines	101 nt	97%	1.21×
Arabidopsis Tissues	101 nt	95%	1.12×

Table 2: Comparative Analysis of 1-pass vs. 2-pass Alignment in Splicing Analysis [7] A study using the MAJIQ pipeline to analyze differential splicing in GTEx and ENCODE data revealed trade-offs between discovery and reproducibility.

Metric	1-Pass Alignment	2-Pass Alignment (Unfiltered)	2-Pass Alignment (Filtered)
Splicing Changes Detected	Baseline	More LSVs found	More LSVs found than 1-pass
Uniquely Mapped Reads	Baseline	1-2% decrease	~0.4% decrease
Run Time	Baseline	3-5 min/sample increase	1-2 min/sample increase
Reproducibility of Unique LSVs	Higher for 1-pass-only LSVs	Lower for 2-pass-only LSVs	Still lower than 1-pass-only LSVs
Recommendation	Preferred for most studies	Useful for hypothesis-generation	Mitigates downsides of unfiltered 2-pass

Table 3: Effect of Read Length and Depth on Junction Calling Accuracy [22] Simulated data analysis reveals fundamental trends affecting all RNA-seq mappers, including STAR.

Experimental Condition	Effect on Splice Junction Recall	Effect on Splice Junction Precision
Increasing Read Length (e.g., 76bp to 101bp)	Improves	Improves
Increasing Sequencing Depth	Marginally Improves	Significantly Decreases

The Scientist's Toolkit

This table lists essential reagents, software, and data resources required for implementing the protocols described in this note.

Item	Function / Description	Example / Source
STAR Aligner	Ultra-fast RNA-seq read aligner capable of detecting canonical and non-canonical splice junctions.	GitHub Repository [14]
Reference Genome	The genomic sequence for the organism of interest.	GRCh38 for human, GRCm39 for mouse (ENSEMBL, Gencode)
Gene Annotation File	A file in GTF or GFF3 format containing known gene models and splice junctions.	Gencode Basic annotations are recommended to exclude poorly supported transcripts [8].
Portcullis	A tool for rapidly and accurately filtering false-positive splice junctions from RNA-seq BAM files.	GitHub Repository [22]
Splice Junction Database	A file listing high-confidence splice junctions, used to guide the second pass of alignment.	Typically generated from a first pass of STAR (SJ.out.tab) and optionally filtered.
Trim Galore	A wrapper tool for Cutadapt and FastQC that performs automated adapter and quality trimming.	Used for pre-processing in published splicing studies [23].
Gynosaponin I	Gynosaponin I, MF:C42H72O12, MW:769.0 g/mol	Chemical Reagent
Autogramin-2	Autogramin-2, MF:C21H27N3O4S, MW:417.5 g/mol	Chemical Reagent

The accuracy of differential splicing analysis is fundamentally constrained by the quality of the input alignment files. Preparing BAM files that optimally represent splice junctions requires careful consideration of alignment strategies and parameters, particularly when using the popular STAR aligner. Within a broader research thesis on STAR alignment parameters for novel splice junction detection, this protocol details the steps for generating and refining BAM files to ensure they are properly structured for downstream differential splicing tools such as rMATS, DEXSeq, and Bisbee. We present a standardized workflow, benchmarked performance data, and reagent solutions to facilitate robust splicing analysis in research and drug development contexts.

Alignment Strategy Comparison: One-Pass vs. Two-Pass STAR

The choice between one-pass and two-pass alignment with STAR significantly impacts splice junction detection sensitivity, especially for novel (unannotated) junctions. The following section quantitatively compares these approaches.

Table 1: Performance comparison of one-pass versus two-pass STAR alignment

Performance Metric	One-Pass Alignment	Two-Pass Alignment	Technical Rationale
Novel Junction Quantification	Baseline	Up to 1.7-fold median read depth improvement [8]	Uses junctions discovered in Pass 1 as annotations for Pass 2, reducing bias against novel junctions.
Sensitivity	Standard	Higher sensitivity for novel and low-coverage junctions [8]	Lower stringency in the second pass allows alignment of reads with shorter overhangs.
Computational Load	Lower (Baseline)	Higher (3-5 minutes more per sample) [7]	Requires two sequential alignment steps and generation of a new genome index.
Unique Read Mapping Rate	Higher (Baseline)	0.4% - 2% reduction [7]	Increased number of annotated junctions can lead to more multi-mapping reads.
Reproducibility of Splicing Changes	High for core events	Detects more LSVs, but additional events can be less reproducible [7]	Junctions with low support from the first pass may introduce less reliable signals.

The core principle of two-pass alignment is to enhance sensitivity. In the first pass, splice junctions are discovered de novo with high stringency. These discovered junctions are then used as a custom annotation file during the second alignment pass, allowing the aligner to recognize them with the same sensitivity as pre-defined annotated junctions [8].

The following diagram illustrates the logical workflow and decision process for choosing and implementing an alignment strategy:

Experimental Protocol for BAM File Preparation

This section provides a detailed, step-by-step methodology for generating high-quality BAM files suitable for differential splicing analysis.

Protocol 1: Two-Pass STAR Alignment with Junction Filtering

Primary Objective: To generate a BAM file with enhanced sensitivity for novel splice junctions while controlling for potential false positives.

Materials and Reagents:

• Computational Resources: High-performance computing cluster recommended.
• Software: STAR aligner (v2.4.0h1 or newer), SAMtools.
• Data Input: RNA-seq FASTQ files (paired-end recommended), reference genome (e.g., GRCh38), high-quality gene annotation in GTF format (e.g., GENCODE).

Methodology:

• First Pass Alignment & Junction Discovery:
  • Align RNA-seq reads using STAR with standard parameters and high stringency for novel junction detection.
  • Key Parameters: --alignSJoverhangMin 8 (require 8 nt overhang for novel junctions), --alignIntronMin 20, --alignMatesGapMax 1000000 [8].
  • The critical output is the SJ.out.tab file, which contains all detected splice junctions.

• Filtering Discovered Junctions (Critical Step):

  • Concatenate SJ.out.tab files from multiple samples if applicable.
  • Filter the combined junction list to remove low-quality artifacts. Retain only junctions that are:
    • Canonical: Filter out non-canonical splice sites (column 5 != 0 in SJ.out.tab).
    • High-Coverage: Remove junctions with low read support (e.g., column 7 < 5) [7].
    • Non-Mitochondrial: Exclude mitochondrial junctions (chromosome "chrM") to reduce noise.

• Second Pass Alignment:

  • Use the filtered junction list from Step 2 as the --sjdbFileChrStartEnd input for STAR.
  • Perform the second alignment pass using the original FASTQ files and the same reference genome. This pass will treat the filtered, discovered junctions as "annotated," significantly improving the mapping rate of reads spanning novel junctions.

Protocol 2: Splice Junction Validation with Portcullis

Primary Objective: To further refine the BAM file by culling alignments associated with invalid splice junctions, producing a cleaner resource for downstream analysis [24].

Materials and Reagents:

• Software: Portcullis (available via Docker, Singularity, Conda, or source).
• Input: A sorted BAM file and its index (.bai) from Protocol 1, and a reference genome in FASTA format.

Methodology:

• Install Portcullis: The most straightforward method is via Conda: conda install portcullis -c bioconda [24].
• Run the Full Portcullis Pipeline: The full mode executes all necessary steps in sequence.

• Outputs: Portcullis generates a filtered junction file and, optionally, a filtered BAM file (*.portcullis.bam) where reads supporting invalid junctions have been removed. This refined BAM is ideal for tools that perform their own junction counting.

Downstream Integration with Differential Splicing Tools

Properly prepared BAM files are the starting point for various differential splicing algorithms. The choice of tool depends on the specific biological question and the type of splicing events of interest.

Table 2: Differential splicing tools and their input requirements

Tool	Statistical Method	Primary Input	Splicing Events Detected	Key Consideration
rMATS-turbo [25] [26]	Replicate Multivariate Analysis	BAM or directly from FASTQ	SE, A5SS, A3SS, MXE, RI	Can be resource-heavy; requires consistent read lengths if using BAM.
Bisbee [27]	Beta-Binomial Model	Splice event counts (e.g., from SplAdder)	Annotated and novel events with protein-effect prediction	Focuses on Percent Spliced In (PSI); integrates proteomic validation.
DEXSeq [28]	Negative Binomial GLM (DESeq2-based)	HTSeq-count-like exon bin counts	Differential exon usage	Works on "exon bins," providing a gene-level perspective on usage.
LeafCutter [27]	Non-parametric	Intron excision counts	Splicing clusters (intron retention)	Detects variation in intron usage without pre-defined event types.

The Scientist's Toolkit: Essential Research Reagents & Software

A curated list of key materials and computational tools required for implementing the protocols described in this application note.

Table 3: Research Reagent Solutions for Splicing Analysis

Item Name	Function / Purpose	Specifications / Notes
STAR Aligner [16]	Spliced alignment of RNA-seq reads to a reference genome.	Fast; supports splice-junction and fusion detection; requires a genome index.
Portcullis [24]	Filters false positive splice junctions from RNA-seq alignment BAM files.	Improves junction quality for downstream analysis; outputs filtered BAM/junctions.
rMATS-turbo [26]	Detects differential alternative splicing from RNA-seq data.	Analyzes five core event types; efficient for large-scale datasets.
GENCODE Annotation	A high-quality reference gene annotation.	Provides a comprehensive set of known splice junctions for alignment guidance.
SAMtools	A suite of utilities for processing and viewing alignments.	Used for sorting, indexing, and manipulating BAM files; a foundational tool.

The integration between alignment preparation and downstream splicing analysis is critical for generating biologically meaningful results. The application of a two-pass STAR alignment strategy, followed by rigorous junction filtering with tools like Portcullis, produces BAM files of sufficient quality to empower differential splicing detection tools like rMATS and Bisbee. While two-pass alignment enhances novel junction discovery, researchers must balance sensitivity with reproducibility and computational overhead. The protocols and benchmarks provided here offer a reliable roadmap for researchers and drug development professionals to standardize their splicing analysis workflows, thereby increasing the robustness of conclusions drawn in both basic research and clinical applications.

The accurate detection of novel splice junctions from RNA-seq data is a cornerstone of advanced transcriptomics, with significant implications for understanding gene regulation, disease mechanisms, and drug target discovery. This process is technically challenging due to the inherent limitations of alignment algorithms, which must map short sequencing reads to reference genomes while distinguishing canonical splicing events from biological novelties and technical artifacts. The core challenge lies in the algorithmic bias of aligners toward known, annotated junctions, which can suppress the detection of unannotated splicing events [8].

The choice of alignment parameters and strategies directly influences detection sensitivity, specificity, and ultimately, the reproducibility of downstream analyses. This case study examines the implementation of a reproducible pipeline for novel splice junction detection, framed within a broader investigation of Spliced Transcripts Alignment to a Reference (STAR) parameters. We focus specifically on the empirical comparison of one-pass versus two-pass alignment modes, a critical methodological decision point that significantly impacts novel junction quantification [8] [7].

Key Concepts and Algorithmic Foundations

The STAR Alignment Algorithm

STAR employs a novel strategy for spliced alignments based on sequential maximum mappable seed search in uncompressed suffix arrays followed by seed clustering and stitching. This two-phase approach represents a fundamental departure from earlier algorithms that were extensions of DNA short-read mappers [6].

• Seed Search Phase: The algorithm identifies the Maximal Mappable Prefix (MMP), defined as the longest substring from the read start that matches one or more genomic locations exactly. When a splice junction or misalignment is encountered, the MMP search restarts from the first unmapped base of the read, naturally identifying splice junction boundaries in a single alignment pass without prior knowledge of junction loci [6].
• Clustering and Stitching Phase: All seeds aligned in the first phase are clustered by genomic proximity and stitched together using a dynamic programming algorithm that allows for mismatches and small indels. This principled approach allows STAR to handle non-canonical splices and chimeric transcripts while maintaining exceptional mapping speed [6].

The Two-Pass Alignment Rationale

Standard single-pass alignment gives preference to known splice junctions, which reduces noise but introduces systematic bias against novel junctions by requiring stronger evidence for their detection [8]. Two-pass alignment addresses this limitation through a elegant strategy:

• First Pass: High-stringency alignment is performed to discover splice junctions de novo from the data.
• Second Pass: The junctions discovered in the first pass are supplied as an augmented annotation file, enabling the aligner to apply lower stringency parameters specifically for these newly documented junctions, thereby increasing mapping sensitivity [8].

Experimental Protocols and Methodologies

• Alignment Software: STAR (version 2.4.0h1 or later) for its speed, sensitivity, and fine-grained parameter control [8].
• Reference Genomes: Use standard references (e.g., GRCh38 for human, TAIR10 for Arabidopsis) to ensure consistency with public data resources [8].
• Gene Annotation: For human studies, the GENCODE-Basic gene annotation set (e.g., v21) provides a comprehensive yet high-quality transcriptome reference excluding poorly supported nominations [8].
• Computing Infrastructure: A modest 12-core server suffices for processing large datasets (e.g., 550 million paired-end reads per hour) [6].

Core Two-Pass Alignment Protocol with Filtering

Step 1: First-Pass Alignment and Junction Discovery

This initial pass generates a SJ.out.tab file containing all discovered splice junctions [7].

Step 2: Junction Filtering Critical for managing computational burden and reducing false positives, this step processes the SJ.out.tab file by:

• Removing junctions with read count < 5 (column 7) [7]
• Excluding non-canonical junctions (column 5 = 0) [7]
• Filtering out mitochondrial genes (column 1 = "chrM") [7]

Step 3: Second-Pass Alignment

The filtered junction file from Step 2 is supplied via the --sjdbFileChrStartEnd parameter, creating a sample-specific augmented reference for more sensitive alignment [8] [7].

Parameter Optimization for Splice Junction Detection

Beyond the basic workflow, these STAR parameters critically affect junction detection performance:

• --alignSJoverhangMin 8: Requires reads span novel splice junctions by at least 8 nucleotides for specificity [8].
• --alignSJDBoverhangMin 3: Requires reads span known splice junctions by at least 3 nucleotides, balancing sensitivity and error reduction [8].
• --alignIntronMin 20 and --alignIntronMax 1000000: Sets biologically plausible intron size boundaries [8].
• --scoreGenomicLengthLog2scale 0: Eliminates intron length-based scoring penalties that can bias against long introns [8].

Experimental Validation Protocol

While computational detection is essential, experimental validation remains crucial for confirming novel junctions:

• RT-PCR Amplification: Design primers flanking predicted novel junctions using RNA from the original sample [6].
• Roche 454 Sequencing: Sequence RT-PCR amplicons to validate junction structure at nucleotide resolution [6].
• Validation Rate Assessment: Calculate the percentage of computationally predicted junctions confirmed experimentally. Published validation studies report success rates of 80-90% for novel intergenic junctions detected by STAR [6].

Results and Performance Benchmarking

Quantitative Assessment of Two-Pass Alignment Benefits

Table 1: Performance Improvements with Two-Pass Alignment Across Diverse RNA-seq Datasets

Sample Type	Description	Read Length	Junctions Improved	Median Read Depth Ratio
Lung Adenocarcinoma	Tumor vs. Normal Tissue	48 nt	98-99%	1.68-1.71×
Reference RNA	Universal Human Reference RNA	75 nt	94-97%	1.25-1.26×
Lung Cancer Cell Lines	Multiple Cell Lines	101 nt	97%	1.19-1.21×
Arabidopsis Tissues	Flower Buds & Leaves	101 nt	95-97%	1.12×

The data demonstrate that two-pass alignment consistently improves quantification across diverse biological contexts, with the most substantial benefits observed in human tumor samples [8].

Trade-offs Between Detection Power and Reproducibility

Independent analyses comparing one-pass versus two-pass alignment reveal critical operational considerations:

• Detection Sensitivity: Two-pass alignment identifies more splicing changes (LSVs) than one-pass across multiple datasets including GTEx tissues and ENCODE knockdown experiments [7].
• Reproducibility Cost: The additional LSVs detected only by two-pass alignment show lower reproducibility across biological replicates compared to events detected by both methods [7].
• Computational Overhead: Two-pass mapping increases runtime (3-5 minutes per sample) and decreases uniquely mapped reads by 1-2%, effects that amplify in larger studies [7].

Impact on Splicing Quantification Accuracy

Table 2: Comparative Analysis of One-Pass vs. Two-Pass Alignment Outcomes

Performance Metric	One-Pass Alignment	Two-Pass Alignment
Number of Detected Splicing Changes	Baseline	Increased
Reproducibility of Findings	Higher	Lower for uniquely detected events
Computational Requirements	Lower	Higher (time + memory)
dPSI Correlation Between Methods	Reference	High (>0.99) for shared events
Effect on Novel Junction Quantification	Under-quantification	Improved quantification

While two-pass alignment improves sensitivity, the dPSI values (change in percent-spliced-in) for most events show minimal differences between methods, with ~99% of changes < 0.025, indicating generally consistent quantification for the majority of splicing events [7].

The Scientist's Toolkit

Table 3: Essential Research Reagents and Computational Resources

Resource	Function/Application	Specifications/Considerations
STAR Aligner	Spliced alignment of RNA-seq reads	Version 2.4.0h1+; requires C++ compilation [8]
GENCODE Annotation	Reference transcriptome for alignment	Use "Basic" gene set for balanced sensitivity/specificity [8]
GRCh38 Reference	Primary alignment genome	Use "full" version without alternate contigs [8]
High-Performance Computing	Execution of alignment pipeline	12+ cores; 32GB+ RAM recommended for mammalian genomes [6]
RT-PCR Reagents	Experimental validation of novel junctions	Follow manufacturer's protocols for RNA template [6]

Workflow Visualization

Two-Pass Novel Junction Detection Pipeline

Decision Framework for Alignment Strategy Selection

The implementation of a reproducible novel junction detection pipeline requires careful consideration of the trade-offs between detection sensitivity and result reliability. While two-pass alignment with STAR demonstrably improves quantification of novel splice junctions—providing as much as 1.7-fold deeper read coverage—this comes with operational costs including reduced reproducibility of uniquely detected events and increased computational requirements [8] [7].

For most focused studies where specific splicing events are of interest, one-pass alignment provides sufficient sensitivity with superior reproducibility and faster computation. For broad, discovery-oriented investigations where comprehensive junction detection is prioritized, two-pass alignment with stringent junction filtering represents the optimal approach, particularly when supplemented with experimental validation [7].

The reproducibility of any novel junction detection pipeline depends critically on comprehensive documentation of parameters, version control for all software components, and transparent reporting of filtering strategies. By implementing the protocols and considerations outlined in this case study, researchers can establish robust, reproducible workflows for splice junction detection that advance our understanding of transcriptome complexity and its implications for disease and drug development.

Solving Common Challenges: Balancing Sensitivity, Specificity, and Computational Efficiency

Large-scale RNA-seq studies, such as those involving thousands of samples, present significant computational challenges. The STAR (Spliced Transcripts Alignment to a Reference) aligner is a cornerstone for splice junction detection but requires substantial system resources and careful parameter configuration. Genome indexing, in particular, is a memory-intensive process demanding at least 32GB of RAM for optimal performance with large genomes [29]. As dataset scale increases, strategies to manage computational burden—such as optimizing alignment passes and filtering junction data—become critical for efficient and accurate novel splice junction discovery. This document outlines proven strategies to address these challenges.

Optimization Strategies and Comparative Analysis

Selecting appropriate alignment strategies and parameters is fundamental to balancing computational cost with result accuracy. The choice between one-pass and two-pass alignment, along with specific filtering thresholds, directly impacts runtime, mapping rates, and the reliability of downstream splicing analysis.

Table 1: Comparison of STAR Alignment Strategies for Large-Scale Studies

Strategy	Key Parameters & Actions	Impact on Performance	Impact on Splice Junction Detection
One-Pass Alignment	Single alignment to the reference genome.	Faster (saves 3-5 min/sample) Higher uniquely mapped reads (by 1-2%) [7]	Fewer novel junctions detected; high reproducibility for detected events [7].
Two-Pass Alignment	First pass discovers junctions; second pass uses these as annotations.	Slower Lower uniquely mapped reads [7]	Detects more potential novel junctions, but a subset may be less reproducible [7].
Filtered Two-Pass	Filter SJ.out.tab from first pass: remove low-coverage (e.g., < 5 reads), non-canonical junctions, and mitochondrial junctions [7].	Moderate runtime increase (1-2 min/sample) Small drop in unique reads (0.4%) [7]	Increases overlap with one-pass results; reduces spurious junctions while retaining sensitivity [7].

Experimental Protocol: Implementing a Filtered Two-Pass Workflow

This protocol is designed for processing dozens to hundreds of RNA-seq samples for novel splice junction discovery, balancing thoroughness with computational efficiency.

1. Resource Provisioning and Genome Indexing

• System Requirements: Ensure access to a 64-bit Linux system with a minimum of 8 CPU cores (16+ ideal) and at least 32GB of RAM [29].
• Genome Index Generation: Generate the STAR genome index using comprehensive reference genome (FASTA) and annotation (GTF) files [29].

2. First-Pass Alignment and Junction Extraction

• Execute the first alignment pass for all samples. This step maps reads and generates a file of detected splice junctions for each sample (SJ.out.tab).

3. Junction Filtering and Consolidation

• Concatenate all SJ.out.tab files and apply filters to create a high-confidence, sample-specific annotation file for the second pass.

4. Second-Pass Alignment

• Realign all reads using the filtered splice junction file. This informs the aligner of novel junctions, improving mapping accuracy.

Experimental Validation and Data Analysis

The effectiveness of alignment strategies must be validated by examining their impact on downstream splicing analysis, such as the quantification of alternative splicing events.

Table 2: Impact of Alignment Strategy on Splicing Quantification (MAJIQ Analysis)

Metric	One-Pass vs. Filtered Two-Pass Findings
Overall dPSI Correlation	~99% of splicing events show minimal difference (	dPSI	< 0.025) [7].
Significant Event Overlap	High concordance for events detected by both methods; strong dPSI correlation [7].
Strategy-Specific Events	Each method detects a small subset of significant events unique to it [7].
Reproducibility of Unique Events	Events detected only in two-pass mode tend to be less reproducible across independent sample sets compared to one-pass-only events [7].

Table 3: Key Computational Tools and Reagents for Splice Junction Discovery

Item	Function / Purpose
STAR Aligner	A splice-aware aligner that accurately maps RNA-seq reads to a reference genome, detecting both annotated and novel splice junctions [29].
Reference Genome (FASTA)	The genomic sequence for the target organism, required for building the alignment index [29].
Gene Annotation (GTF/GFF)	File containing known gene models and exon boundaries, used to guide the initial genome indexing and alignment [29].
High-Confidence Filtered Junctions	A custom splice junction database, created from empirical data, used to enhance sensitivity in a two-pass alignment protocol [7].
Splicing Quantification Tool (e.g., MAJIQ)	Software that interprets aligned RNA-seq data to identify and quantify alternative splicing events, such as cassette exons or intron retention [7].

Workflow and Decision-Making

The following diagram illustrates the core experimental workflow and the logical decision points for strategy selection.

Decision Workflow for STAR Alignment Strategy

Key Parameter Specifications

Fine-tuning specific parameters allows researchers to optimize performance for their specific computational environment and data type.

Key Parameter Specifications

The detection of novel splice junctions from RNA sequencing (RNA-seq) data is a fundamental step in understanding transcriptome diversity and identifying splicing alterations in disease. However, this process is frequently compromised by false positive junctions arising from alignment artifacts, sequencing errors, and biological contaminants. In the context of research utilizing the STAR aligner, mitigating these false discoveries is paramount for ensuring biological validity. False positives can stem from multiple sources, including misalignment of reads in complex genomic regions, strand-specific artifacts, and the inherent error profiles of long-read sequencing technologies [30] [31] [15]. The following sections provide a detailed protocol for implementing a multi-layered filtering strategy, integrating read-based metrics, alignment parameters, and orthogonal validation to distinguish reliable novel junctions from spurious alignments.

A Multi-Faceted Filtering Strategy

Foundational STAR Alignment Parameters for Initial Quality Control

The first line of defense against false positives involves configuring the STAR aligner with parameters that enforce stringent alignment quality. A standard two-pass method is recommended for sensitive novel junction discovery. The following command illustrates key parameters for the initial alignment pass:

Parameter Rationale:

• --outFilterType BySJout: This critical option reduces the number of spurious junctions by filtering out alignments that contain junctions with poor evidence, right after the initial mapping step [32] [31].
• --outSJfilterCountUniqueMin: Sets the minimum number of uniquely mapping reads required to support a junction. The values are provided for different splice site motifs (e.g., non-canonical, GT/AG, etc.) [30] [31]. A higher threshold for non-canonical motifs is advised.
• --outSJfilterOverhangMin: Defines the minimum overhang (the number of bases aligning to each exon) for a junction. A longer overhang (e.g., 30 bp for non-canonical) significantly increases confidence in the alignment [31].
• --outFilterIntronMotifs RemoveNoncanonical: This removes junctions with non-canonical splice sites (not GT/AG, GC/AG, or AT/AC), which are statistically more likely to be alignment artifacts, though it may sacrifice some rare biological events [30].

Table 1: Key STAR Filtering Parameters for Junction Detection

Parameter	Recommended Setting	Function	Impact on FP Reduction
--outFilterType	BySJout	Filters alignments post-junction-discovery	High
--outSJfilterCountUniqueMin	3 2 2 2	Min unique reads per junction motif type	High
--outSJfilterCountTotalMin	10 5 5 5	Min total reads per junction motif type	Medium
--outSJfilterOverhangMin	30 12 12 12	Min overhang length per junction motif type	High
--outFilterIntronMotifs	RemoveNoncanonical	Removes non-canonical splice sites	Medium-High
--alignIntronMax	200000	Maximum allowed intron size	Low

Post-Alignment Read-Based Filtering and Quantification

After alignment, junctions must be filtered based on the quantitative evidence from the BAM files. This step involves aggregating data across samples to distinguish recurrent technical artifacts from genuine, albeit lowly expressed, junctions.

Protocol: Read-Count Based Filtering

• Extract Junction Read Counts: Use STAR's SJ.out.tab files or a tool like featureCounts from the Rsubread package to compile a table of all detected junctions and their supporting read counts per sample [32].
• Apply Universal Thresholds: Filter the junction list to retain only those supported by a minimum number of uniquely mapping reads. A common starting threshold is 3-5 unique reads per sample. Junctions with exclusively multi-mapped reads should be treated with caution, as they often arise from duplicated genomic regions [30].
• Implement Recurrence Filtering: Require that a junction is present in multiple samples or replicates. A junction found in at least two samples within a condition greatly reduces the likelihood of a random artifact.
• Calculate Proportionate Support: For genes with high multi-mapping activity, calculate the ratio of unique to multi-mapped reads supporting a junction. A low ratio (e.g., <10%) suggests the "unique" reads may be false positives in a duplicated locus [30].

Table 2: Post-Alignment Filtering Metrics for Novel Junctions

Metric	Calculation	Recommended Threshold	Rationale
Unique Read Count	Number of uniquely aligned reads spanning the junction.	>= 3-5	Ensures basic evidence for the junction.
Total Read Count	Unique + multi-mapped reads spanning the junction.	>= 5-10	Provides a measure of overall expression.
Sample Recurrence	Number of samples in which the junction is detected.	>= 2	Filters sample-specific artifacts.
Unique-to-Total Ratio	Unique Count / Total Count	> 0.1	Flags junctions in regions of high ambiguity.

Advanced and Orthogonal Validation Methods

For critical applications, such as clinical diagnostics or validating key findings, more advanced computational and experimental methods are required.

Advanced Computational Filtering with Machine Learning: Tools like FineSplice and 2passtools employ machine learning to identify subtle patterns indicative of false positives.

• FineSplice Protocol: This pipeline uses a semi-supervised anomaly detection method.
  • Perform initial alignment with TopHat2 or STAR.
  • For each junction, compute the set of split-read overhangs.
  • Define a subset of potential false positives based on the distribution of overhangs and the presence of mismatches.
  • Fit a logistic regression model using features derived from the overhang distribution.
  • Discard junctions with a high posterior probability of being false positives [33].
• 2passtools Protocol: This tool is particularly effective for long-read RNA-seq data.
  • Perform a first-pass alignment with a long-read aligner like minimap2.
  • Extract alignment metrics and sequence information from the first-pass junctions.
  • Use a rule-based filter and a logistic regression model to classify and remove spurious junctions.
  • Use the filtered, high-confidence junctions as a guide for a second-pass alignment, which dramatically improves accuracy [15].

Orthogonal Experimental Validation: Computational predictions must be confirmed experimentally.

• RT-PCR and Sanger Sequencing: Design primers in the flanking exons of the putative novel junction. Amplify cDNA from the original RNA sample and sequence the product to confirm the exact splice site.
• Orthogonal Sequencing: Support findings with data from another sequencing platform. For example, short-read Illumina data can be used to validate junctions found in long-read ONT or PacBio data, and vice-versa [34] [9]. The high accuracy of short reads is excellent for confirming junction support.
• Independent RNA-seq Analysis: Process the same RNA sample with a different RNA-seq library preparation method (e.g., CapTrap) or a different alignment tool (e.g., TopHat2, HISAT2) to see if the junction is independently detected [34].

Table 3: Key Research Reagent Solutions for Junction Validation

Item	Function / Application	Example Use Case
STAR Aligner	Spliced alignment of RNA-seq reads to a reference genome.	Primary detection of splice junctions from FASTQ files. [30] [32]
Rsubread/featureCounts	Quantification of read counts over genomic features, including exon junctions.	Generating count matrices for exon-junction usage analysis. [32]
FineSplice	Machine-learning based post-processing of alignments to filter false positive junctions.	Enhancing precision of junction calls from STAR/TopHat2 alignments. [33]
2passtools	Machine-learning filtered two-pass alignment for long-read RNA sequencing.	Improving intron detection accuracy in PacBio and ONT data. [15]
Twist Mouse Exome Panel	Targeted exome capture for long-read sequencing.	Enriching for spliced, coding reads to improve transcriptome complexity. [35]
SIRV-Set 4 Spike-in	A set of synthetic RNA spike-ins with known splice variants.	Benchmarking and optimizing splice junction detection performance. [34]

Visualizing the Validation Workflow

The following diagram illustrates the integrated, multi-stage workflow for novel junction detection and validation, from raw data to high-confidence junctions.

Diagram 1: A multi-phase workflow for validating novel splice junctions, incorporating stringent alignment, quantitative filtering, and orthogonal validation to mitigate false positives.

Validating novel splice junctions is an iterative process that balances sensitivity with specificity. No single filter is sufficient to eliminate all false positives. Instead, a combination of stringent STAR parameters, evidence-based read filtering, and, for high-stakes discoveries, advanced machine learning and orthogonal experimental validation is required. The protocols and strategies outlined herein provide a robust framework for researchers to enhance the reliability of their novel junction predictions, thereby ensuring the biological insights derived from RNA-seq data are built upon a solid foundation. As sequencing technologies and algorithms continue to evolve, particularly for long-read data, these filtering strategies will remain a critical component of the transcriptomic analysis toolkit.

Handling Intron Retention Events and Other Challenging Splicing Patterns

Intron retention (IR) is a form of alternative splicing where introns are deliberately retained in mature mRNAs, contrasting with conventional splicing that removes all introns prior to export and translation [36]. Once dismissed as splicing noise, IR is now recognized as a dynamic and evolutionarily conserved mechanism of post-transcriptional gene regulation that influences mRNA stability, localization, and translational potential [36]. Retained introns can lead to nonsense-mediated decay (NMD), promote nuclear retention, or give rise to novel protein isoforms that contribute to expanding proteomic and transcriptomic profiles [36]. IR plays critical roles in cell-type and tissue-specific gene expression and functions as a molecular switch during cellular responses to environmental stressors such as hypoxia, heat shock, and infection [36]. Dysregulated IR is increasingly associated with cancer, neurodegeneration, aging, and immune dysfunction, where it may alter protein function, suppress tumor suppressor genes, or generate immunogenic neoepitopes [36].

The regulation of IR is a multifactorial process influenced by splice site strength, splicing regulatory elements (SREs), chromatin structure, methylation patterns, RNA polymerase II elongation rates, and the availability of co-transcriptional splicing factors [36]. Retained introns are often shorter, GC-rich, and flanked by weak splice sites, all of which make them less likely to be removed during splicing [36]. RNA-binding proteins (RBPs) including hnRNPLL, PTBP1, and ASF/SF2 play key roles in IR regulation through direct interactions with pre-mRNAs [36].

Wet-Lab Protocols for Capturing Splicing Events

Minimally Invasive RNA-seq Using Peripheral Blood Mononuclear Cells (PBMCs)

For rare disease research and clinical diagnostics, accessible tissues are essential for transcriptomic analysis. A 2025 study established a minimally invasive RNA-seq protocol using short-term cultured peripheral blood mononuclear cells (PBMCs) that enables detection of transcripts subject to nonsense-mediated decay [37]. This protocol is particularly suited for neurodevelopmental disorders, as up to 80% of the genes in intellectual disability and epilepsy gene panels are expressed in PBMCs [37].

Experimental Workflow:

• Sample Collection: Collect peripheral blood samples using standard venipuncture procedures
• PBMC Isolation: Isolate PBMCs using density gradient centrifugation (e.g., Ficoll-Paque)
• Cell Culture: Culture PBMCs for short-term expansion (96 hours) in RPMI-1640 medium supplemented with 10% FBS and 1% penicillin-streptomycin
• NMD Inhibition: Treat aliquots of cells with 100 µg/mL cycloheximide (CHX) or DMSO vehicle control for 4-6 hours before harvesting
• RNA Extraction: Extract total RNA using automated systems (e.g., Promega Maxwell RSC with SimplyRNA Blood kit)
• Quality Control: Assess RNA integrity using TapeStation with RNA ScreenTape (samples should have DV200 >70%)

This protocol demonstrated effectiveness in revealing aberrant splicing in six of nine individuals with splice variants, allowing reclassification of seven variants, and outperformed targeted cDNA analysis in capturing complex splicing events including intron retention [37].

NMD Inhibition and Validation

Nonsense-mediated decay can mask splicing defects by degrading transcripts containing premature termination codons. Effective NMD inhibition is therefore crucial for comprehensive detection of IR events:

Optimized NMD Inhibition Protocol:

• Inhibitor Selection: Use cycloheximide (CHX) at 100 µg/mL rather than puromycin based on superior performance in restoring NMD-sensitive transcripts [37]
• Treatment Duration: Incubate cells for 4-6 hours with CHX before RNA extraction
• Internal Control: Monitor SRSF2 NMD-sensitive transcripts as an endogenous control for NMD inhibition efficacy [37]
• Validation: Confirm inhibition efficiency by quantifying exon 3 spanning reads in SRSF2 transcripts (typically increasing from 4.55% in untreated to 8.58% in CHX-treated samples) [37]

Sample Preparation for Challenging Tissues

When studying disease-relevant tissues that are difficult to access, such as brain in neurodegenerative disorders, alternative approaches are necessary:

Formalin-Fixed Paraffin-Embedded (FFPE) Tissue Protocol:

• RNA Extraction: Use specialized kits for FFPE samples (e.g., Maxwell RSC RNA FFPE kit)
• Quality Assessment: Accept samples with DV200 >70% despite overall RNA degradation typical of FFPE material
• Targeted Enrichment: Employ capture-based RNA sequencing to overcome limited RNA quality [38]

Table 1: Key Research Reagent Solutions for Splicing Analysis

Reagent/Cell Type	Key Function	Application Context	Considerations
PBMCs	Clinically accessible tissue for transcriptomics	Rare disease diagnostics, neurodevelopmental disorders	Express ~80% of ID/epilepsy panel genes; minimal invasiveness [37]
Cycloheximide (CHX)	NMD inhibitor	Revealing PTC-containing transcripts; detecting aberrant splicing	100 µg/mL, 4-6 hour treatment; monitor SRSF2 as internal control [37]
Lymphoblastoid Cell Lines (LCLs)	Immortalized B-cells	Splicing profile characterization; functional validation	Express ~64-75% of disease genes; suitable for CDH1 splicing studies [38]
Fibroblasts	Differentiated connective tissue cells	Broad splicing studies; metabolic disorders	Express ~72% of disease panel genes; more invasive collection [37]

Computational Detection of Splicing Events

Two-Pass Alignment for Novel Junction Discovery

Reduced alignment power traditionally impedes expression quantification of novel splice junctions. Two-pass alignment addresses this limitation by separating splice junction discovery from quantification [8].

Implementation with STAR:

Performance Characteristics: Two-pass alignment significantly improves quantification of novel splice junctions, providing as much as 1.7-fold deeper median read depth over these junctions compared to single-pass alignment [8]. Across diverse RNA-seq datasets, two-pass alignment improved quantification of at least 94% of simulated novel splice junctions [8].

Table 2: Two-Pass Alignment Performance Across Sample Types

Sample Type	Read Length	Splice Junctions Improved	Median Read Depth Ratio	Key Applications
Lung Adenocarcinoma	48-101 nt	96-99%	1.20-1.71×	Cancer splicing variants, novel isoform discovery [8]
Reference RNA (UHRR)	75 nt	94-97%	1.25-1.26×	Method benchmarking, analytical validation [8]
Arabidopsis Tissues	48 nt	95-97%	1.12×	Plant genomics, developmental splicing [8]
Cell Lines	101 nt	97%	1.19-1.21×	Controlled experiments, mechanistic studies [8]

Junction Filtering Strategies

A critical consideration in two-pass alignment is managing the increased number of splice junctions, which can lead to higher computational burden and reduced uniquely mapped reads [7].

Optimized Junction Filtering:

Filtering parameters should exclude:

• Junctions with low coverage (<5 uniquely mapping reads)
• Non-canonical junction types (column 5 = 0)
• Mitochondrial genes (chrM) [7]

This filtering approach reduces the drop in uniquely mapped reads from 1-2% to approximately 0.4% while maintaining detection sensitivity [7].

Advanced Algorithms for Splicing Analysis

Long-Read RNA-seq Approaches: Recent advances in long-read sequencing technologies enable clearer demarcation of cis- and trans-directed splicing events [39]. The isoLASER method leverages long-read RNA-seq to identify allele-specific alternative splicing patterns, distinguishing between:

• Cis-directed events: Characterized by allele-specific alternative splicing patterns linked to heterozygous variants
• Trans-directed events: Exhibit no linkage between haplotypes and splicing, reflecting the dominant role of trans-acting factors [39]

Differential Splicing Analysis with Covariates: Tools like MntJULiP implement sophisticated Bayesian models that adjust for covariates (age, sex, ethnicity), substantially improving accuracy in detecting true biological splicing differences while reducing false positives [40]. The Jutils package provides visualization capabilities through heatmaps, Venn diagrams, and PCA plots for interpreting complex splicing data [40].

Novel Isoform Discovery: Torino is a computational workflow that uses Poisson non-negative matrix factorization with spatial smoothness priors to infer latent transcript structures without pre-existing annotations [41]. Applied to GTEx samples, Torino revealed extensive unannotated diversity, including over 53,000 novel intron retention events, many exhibiting strong tissue specificity [41].

Experimental Design and Workflow Integration

Comprehensive Analytical Workflow

The following diagram outlines an integrated experimental and computational workflow for handling intron retention events:

Strategic Considerations for Alignment Parameter Optimization

Organism-Specific Parameters: When working with non-mammalian systems, default STAR parameters require adjustment. For plant genomes, specific modifications are necessary [42]:

• Intron size range: Adjust to 60-6000 nucleotides (vs. mammalian default 20-1,000,000)
• Splice junction overhang: Potentially reduce from default 8 nucleotides to account for smaller introns

Sensitivity vs. Reproducibility Trade-offs: Empirical evaluations reveal that two-pass alignment detects more splicing changes than one-pass (15-25% additional significant LSVs), but these additional events show lower reproducibility across biological replicates [7]. This suggests:

• Use one-pass alignment for confirmatory studies requiring high reproducibility
• Use two-pass alignment for exploratory, hypothesis-generating studies aiming for maximal sensitivity

Validation Strategies: Computational predictions of splicing defects require experimental confirmation:

• RT-PCR with gel electrophoresis for simple exon skipping events
• Capillary electrophoresis for quantitative fragment analysis
• Digital droplet PCR for precise quantification of specific isoforms [38]

Handling intron retention events and other challenging splicing patterns requires integrated wet-lab and computational approaches. The protocols outlined here—from PBMC processing with NMD inhibition to optimized two-pass alignment with junction filtering—provide a comprehensive framework for detecting and quantifying these biologically significant but technically challenging events. As splicing analysis continues to evolve, emerging technologies like long-read sequencing and advanced computational methods like Torino will further enhance our ability to decipher the complex landscape of alternative splicing in health and disease.

In the field of genomics, establishing robust benchmarks is fundamental for accurately detecting novel biological events, such as splice junctions. This is particularly critical in splice junction detection research, where the choice of alignment parameters and analytical workflows directly impacts the sensitivity and reliability of downstream analyses. This article provides detailed application notes and protocols for benchmarking performance in the context of novel splice junction detection, with a specific focus on the STAR aligner.

Quantitative Benchmarks for Splicing Analysis Workflows

Defining expected performance thresholds is a key step in benchmarking any bioinformatics pipeline. The table below summarizes the False Discovery Rate (FDR) and Statistical Power for two implementations of the Differential Exon-Junction Usage (DEJU) workflow, a method that enhances splicing detection by incorporating exon-exon junction reads [43] [32]. These benchmarks were derived from comprehensive simulation studies under a nominal FDR control of 0.05.

Table 1: Performance of DEJU Workflows Across Different Splicing Patterns and Sample Sizes

Splicing Pattern	Sample Size (n)	DEJU-edgeR		DEJU-limma
		FDR	Power	FDR	Power
Exon Skipping (ES)	3	0.022	0.977	0.043	0.975
	5	0.029	0.991	0.044	0.990
	10	0.038	0.992	0.051	0.992
Mutually Exclusive Exons (MXE)	3	0.030	0.990	0.061	0.991
	5	0.040	0.993	0.062	0.995
	10	0.045	0.995	0.063	0.995
Alternative 3'/5' Splice Site (ASS)	3	0.027	0.839	0.038	0.877
	5	0.027	0.927	0.037	0.947
	10	0.038	0.977	0.047	0.979
Intron Retention (IR)	3	0.030	0.866	0.042	0.880
	5	0.031	0.934	0.041	0.940
	10	0.042	0.964	0.050	0.968

These benchmarks reveal several critical insights for establishing quality thresholds:

• DEJU-edgeR consistently controls the FDR at or below the nominal rate across all splicing patterns, making it a more conservative and reliable choice.
• Detection Power is generally high for all patterns but increases substantially with sample size for more complex events like ASS and IR.
• Intron Retention (IR) events were detectable exclusively by workflows that incorporate junction reads, highlighting a fundamental limitation of older methods [32].

Experimental Protocols for Benchmarking

Protocol: The DEJU Analysis Workflow for Differential Splicing

The DEJU workflow is designed for differential splicing analysis in short-read bulk RNA-seq experiments and integrates exon-junction reads to resolve the double-counting problem inherent in standard exon-level analyses [43] [32].

1. Read Alignment and Junction Discovery

• Tool: STAR aligner (2-pass mapping mode) [43] [32].
• Procedure:
  • Perform first-pass mapping on all samples. Collate the junctions detected across all samples and use this combined set to re-index the reference genome.
  • Execute a second-pass mapping using the re-indexed genome. Use the --outFilterType BySJout option to retain only junctions that pass a filtering threshold (e.g., more than three uniquely mapping reads across all samples) in the final BAM files. This step maximizes sensitivity for novel junction detection.

2. Feature Quantification

• Tool: featureCounts function from the Rsubread package [43] [32].
• Critical Parameters: Set useMetaFeatures = FALSE, nonSplitOnly = TRUE, and juncCounts = TRUE.
• Procedure:
  • Run featureCounts with the above parameters on the BAM files from step 1. This generates two count matrices: one for internal exon reads and another for exon-exon junction reads.
  • Concatenate these two matrices into a single exon-junction count matrix, where each read is uniquely assigned to a single feature (exon or junction).

3. Downstream Differential Splicing Analysis

• Tools: Either diffSpliceDGE function in edgeR or the diffSplice function in limma [43] [32].
• Procedure:
  • Filter lowly expressed exons and junctions using the filterByExpr function in edgeR.
  • Normalize the concatenated count matrix using the Trimmed Mean of M-values (TMM) method with the normLibSizes function.
  • Perform differential exon-junction usage analysis. The feature-level test results (for each exon and junction) are then summarized at the gene level using the Simes method (to combine feature-level p-values) or an F-test.

Protocol: Detecting Splice-Site Creating Variants (SSCVs) with Juncmut

The juncmut software identifies genomic variants that create novel splice-sites (Splice-Site Creating Variants, or SSCVs) using transcriptome data alone, providing a method to find previously overlooked pathogenic mutations [44].

1. Identification of Rare Splicing Junctions

• Procedure:
  • Extract a set of rare splicing junctions from the transcriptome data. These are defined as junctions that are infrequently observed in large reference datasets (like GTEx) and that span from an annotated exon-intron boundary to an unannotated boundary (primary novel splice-site).

2. Candidate SSCV Generation and Filtering

• Procedure:
  • For each rare junction, list all possible genomic variants that would make the sequence at the primary novel splice-site closer to the consensus donor (AG|GTRAGT) or acceptor (YYNYAG|R) motif compared to the reference genome.
  • Check if any of these candidate variants are observed as mismatch bases in the transcriptome short reads.
  • Realign all short reads around the candidate SSCV to both the predicted abnormal transcript and the normal transcript sequence to rule out alignment artifacts.
  • Apply stringent filters, including the removal of variants with an allele frequency >0.01 in population databases (e.g., gnomAD).

3. Validation and Evaluation

• Procedure:
  • Validate the genomic status of SSCVs using matched whole-genome sequencing data when available. In a benchmark using 1000 Genomes Project data, 152 out of 153 SSCVs identified by juncmut were confirmed as true genomic variants [44].
  • Quantify the splicing impact by calculating the difference in abnormal splicing ratios between SSCV-positive and SSCV-negative samples in independent datasets (e.g., GTEx). For 23 out of 24 distinct SSCVs, this analysis yielded highly significant p-values (≤ 10⁻⁸), confirming a strong biological effect [44].

Workflow Visualization

The following diagrams illustrate the core workflows described in the protocols, providing a visual guide for implementation.

DEJU Splicing Analysis Workflow

Juncmut SSCV Detection Logic

Successful execution of the aforementioned protocols requires a suite of specialized computational tools and data resources. The following table details these essential components.

Table 2: Key Research Reagents and Resources for Splice Junction Benchmarking

Item Name	Type	Primary Function	Key Features/Application
STAR Aligner [43] [32]	Software	Splice-aware alignment of RNA-seq reads.	2-pass mapping mode maximizes novel junction discovery; --outFilterType BySJout filters low-quality junctions.
Rsubread/featureCounts [43] [32]	Software	Quantification of exon and exon-exon junction reads.	Resolves double-counting by generating unique counts for exons and junctions via nonSplitOnly and juncCounts parameters.
edgeR / limma [43] [32]	R Package	Differential splicing analysis.	diffSpliceDGE (edgeR) and diffSplice (limma) functions test for differential usage of exons and junctions.
Juncmut [44]	Software	Detection of splice-site creating variants (SSCVs).	Identifies SSCVs from transcriptome data alone; fine-tuned to remove false-positives.
SpliceAI [44] [45]	AI Model	Prediction of splice-altering effects from nucleotide sequence.	Scores variants for donor/acceptor gain/loss; used in frameworks like SpliPath for rare variant clustering.
SG-NEx Dataset [46]	Data Resource	Benchmarking long-read RNA-seq protocols.	Provides a comprehensive resource for evaluating RNA-seq methods for transcript-level analysis, including isoform detection.
GTEx Dataset [44]	Data Resource	Reference for "normal" splicing.	Provides a large catalog of common splicing patterns and variants across human tissues, used to define rare/aberrant events.
SSCV DB [44]	Database	Registry of Splice-Site Creating Variants.	A valuable resource for discovering novel biological mechanisms and targets for therapeutic intervention.

Validating Detection Accuracy: Benchmarking Against Ground Truth and Complementary Methods

The detection of novel splice junctions represents a pivotal step in advancing our understanding of transcriptomic diversity and its implications in development, cellular identity, and disease. While modern alignment tools, particularly those involving STAR parameters, have significantly enhanced our ability to identify putative novel splicing events, the inherent challenges of sequencing technologies and algorithmic limitations necessitate rigorous orthogonal validation. High-throughput sequencing technologies introduce multiple sources of potential artifacts, including spurious alignments due to random sequence matches, sample-reference genome discordance, and technical noise from library preparation. Furthermore, as evidenced by the LRGASP consortium, there exists substantial variability in transcript detection across different bioinformatics pipelines, with moderate agreement among tools and little overlap in transcripts identified by any two pipelines [34]. This variability underscores the critical importance of implementing robust orthogonal methods to distinguish biologically relevant novel splice junctions from technical artifacts, thereby ensuring the reliability of subsequent biological interpretations and their translation into therapeutic applications.

Orthogonal Validation Methodologies: A Comparative Framework

Integration of Short-Read and Long-Read Sequencing Technologies

The convergence of short-read and long-read RNA sequencing data provides a powerful framework for validating novel splice junctions. Short-read RNA-seq, while limited in its ability to resolve full-length transcripts, generates high coverage data that can robustly confirm the existence of specific exon-exon junctions identified through long-read approaches. The LRGASP consortium demonstrated that incorporating orthogonal short-read data significantly improves confidence in transcript models, with many pipelines achieving a high percentage of known transcripts with full support at transcription start sites (TSSs), transcription termination sites (TTSs), and junctions [34]. This multi-platform approach leverages the respective strengths of each technology: the high accuracy and depth of short-read sequencing for junction confirmation, and the transcript-length context provided by long-read sequencing.

Comparative analyses across platforms reveal distinct performance characteristics. PCR-amplified cDNA sequencing with Nanopore generates the highest throughput, while PacBio IsoSeq produces the longest reads on average [46]. Direct RNA sequencing avoids reverse transcription and amplification biases but exhibits 3'-end coverage bias due to sequencing initiation at the poly(A) tail [46]. When validating novel junctions, the choice of platform should be guided by the specific biological question, with platform-aware interpretation of supporting evidence.

Two-Pass Alignment and Its Impact on Detection Accuracy

The implementation of two-pass alignment in STAR represents a significant methodological advancement for novel splice junction discovery and validation. This approach separates the processes of splice junction discovery and quantification, with junctions identified in an initial high-stringency alignment pass subsequently used as annotations in a second, more sensitive alignment pass [8]. Empirical evidence demonstrates that two-pass alignment improves quantification of novel splice junctions, providing as much as 1.7-fold deeper median read depth compared to single-pass approaches [8]. This enhanced sensitivity is particularly valuable for detecting low-abundance splicing events that might otherwise be missed.

However, the application of two-pass alignment requires careful consideration of potential drawbacks. Implementation can decrease the percentage of uniquely mapped reads by 1-2% and substantially increase computational runtime [7]. Furthermore, a systematic evaluation revealed that while two-pass alignment identifies more splicing changes, these additional local splicing variations (LSVs) may be less reproducible than those detected by single-pass alignment [7]. To mitigate these issues, filtering of splice junction annotations between passes—removing junctions with low coverage (< 5 reads), non-canonical splicing motifs, and mitochondrial genes—reduces negative impacts on runtime and mapping specificity without significantly compromising sensitivity [7].

Table 1: Performance Comparison of Alignment Strategies for Novel Junction Detection

Alignment Strategy	Sensitivity for Novel Junctions	Quantification Accuracy	Computational Demand	Key Applications
STAR Single-Pass	Baseline	Moderate	Lower	Standard transcriptome quantification
STAR Two-Pass (Unfiltered)	High	Improved	Higher	Comprehensive novel junction discovery
STAR Two-Pass (Filtered)	High	Improved	Moderate	Large-scale studies with reproducibility
Targeted Enrichment (LSV-seq)	Highest for targeted events	High for targeted events	Variable	Hypothesis-driven validation

Computational and Machine Learning Approaches

Advanced computational methods provide powerful orthogonal validation without additional wet-lab experimentation. Machine learning classifiers, such as DeepSplice, leverage convolutional neural networks to distinguish true biological splice junctions from artifacts based on sequence features and alignment characteristics [47]. This approach treats donor and acceptor sites as functional pairs, capturing remote relationships between features that determine splicing outcomes. When applied to a benchmark dataset, DeepSplice outperformed state-of-the-art methods, achieving superior sensitivity and specificity for both donor and acceptor site classification [47].

The SICILIAN (SIngle Cell precIse spLice estImAtioN) framework implements a statistical approach that assigns confidence scores to splice junctions based on multiple alignment features, including the number of alignments per read, read overhang lengths, alignment scores, mismatch counts, soft-clipped bases, and read entropy [48]. This method is particularly valuable for single-cell RNA-seq data, where technical noise is amplified. SICILIAN significantly improves concordance between matched single-cell and bulk datasets, increasing the fraction of junctions detected in single cells that are also present in bulk data from the same cell line from 0.54 to 0.75 in one evaluation [48].

Table 2: Computational Tools for Splice Junction Validation

Tool	Methodology	Key Features	Performance Metrics
DeepSplice	Convolutional Neural Networks	Models donor-acceptor pairs; Uses flanking sequences	auROC: 0.983 (donor), 0.974 (acceptor) on HS3D [47]
SICILIAN	Penalized Generalized Linear Model	Incorporates read entropy; Adapts to batch effects	Increases junction call concordance to 0.75 vs 0.54 with STAR alone [48]
Optimal Prime	Machine Learning Primer Design	Optimizes targeted RNA-seq primers; High on-target efficiency	Enables high-throughput splicing quantification with lower sequencing depth [49]

Targeted RNA-Seq for Sensitive Validation

Targeted RNA-seq methods represent a highly sensitive orthogonal approach for validating specific splicing events of interest. LSV-seq (Local Splicing Variation sequencing) utilizes multiplexed reverse transcription from pools of primers anchored near splicing events to enrich for junction-spanning reads [49]. This method significantly increases on-target capture rates compared to standard RNA-seq, enabling precise quantification of splicing changes with substantially lower sequencing depth. The machine learning algorithm Optimal Prime further enhances this approach by optimizing primer design based on performance data from thousands of primer sequences, achieving high on-target efficiency and enabling the discovery of hundreds of tissue-specific splicing events previously missed due to poor coverage in standard RNA-seq [49].

In hematologic malignancies, targeted RNA-seq panels have demonstrated enhanced detection of splice-altering variants, increasing diagnostic yield compared to DNA gene panel sequencing alone [9]. These approaches efficiently filter out inconsequential splice events generated by deep RNA-seq, focusing attention on clinically significant splice-altering somatic variants with implications for treatment risk assessment and therapeutic decisions [9].

Experimental Protocols and Workflows

Integrated Validation Protocol for Novel Splice Junctions

Diagram 1: Orthogonal validation workflow for novel splice junctions.

Two-Pass Alignment with Junction Filtering Protocol

Procedure:

• First Pass Alignment: Execute STAR alignment with standard parameters to generate initial splice junction predictions.
  • Input: FASTQ files, reference genome, gene annotation (optional)
  • Output: SJ.out.tab file containing discovered junctions

• Junction Filtering: Apply stringent filters to generate high-confidence junction annotations for the second pass.

  • Filter criteria:
    • Remove junctions with read count < 5 (column 7 < 5 in SJ.out.tab)
    • Remove non-canonical junctions (column 5 = 0)
    • Remove mitochondrial junctions (chrM)
  • Concatenate filtered SJ.out.tab files across all samples

• Second Pass Alignment: Realign reads using filtered junctions as annotations.

  • Input: FASTQ files, reference genome, filtered junction annotations
  • Parameters: Increased sensitivity for annotated junctions
  • Output: Final BAM files with improved novel junction quantification [8] [7]

LSV-seq Targeted Enrichment Protocol

Procedure:

• Primer Design: Utilize Optimal Prime machine learning algorithm to design target-specific primers.
  • Design primers to bind downstream of splice junctions of interest
  • Include 10-nucleotide UMI for duplicate removal and error correction

• Library Preparation:

  • Annealing: Gradually anneal LSV-seq primer pool to input RNA using touchdown protocol
  • Reverse Transcription: Perform first-strand cDNA synthesis at 60°C for increased specificity
  • Second Strand Synthesis: Generate double-stranded DNA template
  • In Vitro Transcription: Amplify material using IVT (critical for low-input samples)
  • Fragmentation and Adapter Ligation: Fragment aRNA and append second adapter sequence
  • PCR Amplification: Amplify final library for sequencing [49]

• Data Analysis:

  • Map reads to reference genome
  • Quantify junction coverage for targeted events
  • Compare with standard RNA-seq data for validation

Table 3: Key Research Reagents and Computational Tools for Junction Validation

Category	Specific Tool/Reagent	Function in Validation	Considerations for Use
Sequencing Platforms	Nanopore Direct RNA-seq	Identifies full-length transcripts and native RNA modifications	3'-end coverage bias; No amplification needed [46]
	PacBio IsoSeq	Generates long reads for complex isoform resolution	Depletion of shorter transcripts; Lower throughput [46]
	Illumina Short-read	Provides high-depth confirmation of specific junctions	Fragmentation biases; Limited to short spans [50]
Alignment Tools	STAR Two-Pass	Enhances novel junction discovery and quantification	Increased computational demand; Filtering recommended [8] [7]
Validation Algorithms	DeepSplice	Classifies true vs. false junctions using deep learning	Requires training data; High computational resources [47]
	SICILIAN	Assigns statistical confidence to junction calls	Adapts to batch effects; Incorporates read entropy [48]
Targeted Approaches	LSV-seq Primers	Enriches for specific splicing events of interest	Machine learning-optimized design with Optimal Prime [49]
	Spike-in Controls (Sequins, SIRVs)	Provides quantitative standards for assessment	Platform-specific compatibility (not compatible with direct RNA-seq) [46]
Experimental Validation	RT-qPCR with Reference Genes	Orthogonal confirmation of splicing events	Requires stable, high-expression reference genes (select with GSV software) [51]

The confident identification of novel splice junctions demands an integrated, multi-layered validation strategy that leverages both computational and experimental orthogonal methods. The combination of two-pass alignment with careful junction filtering, computational classification using tools like DeepSplice and SICILIAN, targeted enrichment approaches such as LSV-seq, and confirmation through orthogonal sequencing platforms creates a robust framework for distinguishing true biological splicing events from technical artifacts. This comprehensive approach is particularly crucial in translational research settings, where the accurate detection of splice-altering variants can inform diagnostic and therapeutic decisions. As sequencing technologies continue to evolve and computational methods become increasingly sophisticated, this validation framework will ensure that novel junction discoveries reflect genuine biological phenomena with potential significance for understanding disease mechanisms and developing targeted interventions.

The selection between one-pass and two-pass alignment modes with the STAR (Spliced Transcripts Alignment to a Reference) aligner represents a critical methodological decision in RNA-seq analysis, particularly for research focused on novel splice junction detection. This application note systematically benchmarks the reproducibility of both approaches through quantitative performance metrics. Evidence from controlled analyses of ENCODE and GTEx data indicates that while two-pass alignment increases splice junction detection sensitivity, it introduces supplementary findings with significantly lower reproducibility compared to one-pass modes. We provide detailed experimental protocols and implementation guidelines to assist researchers in selecting the optimal alignment strategy based on their specific research objectives, emphasizing methodological rigor for reliable splicing analysis in both basic research and drug development contexts.

RNA sequencing has become an indispensable tool for exploring transcriptome complexity, with alternative splicing analysis playing an increasingly important role in understanding disease mechanisms and identifying therapeutic targets. The STAR aligner has emerged as a preferred tool for RNA-seq read mapping due to its speed and sensitivity in detecting spliced alignments. However, researchers must choose between one-pass and two-pass alignment modes, each with distinct trade-offs for splice junction detection and quantification [8] [16].

In standard one-pass alignment, STAR utilizes existing gene annotations to guide splice-aware alignment while simultaneously detecting novel splicing events. In contrast, the two-pass approach separates the processes of junction discovery and quantification: the initial alignment pass identifies splice junctions de novo, which are then incorporated as additional "annotations" for a second alignment pass, theoretically improving sensitivity for novel junctions [8]. While this approach has demonstrated benefits for detecting novel splicing events, its impact on measurement reproducibility—a critical requirement for clinical and pharmaceutical applications—remains inadequately characterized.

This application note presents a structured performance comparison between these alignment modes, with particular emphasis on their reproducibility for detecting differential splicing events. We provide quantitative benchmarks derived from real-world datasets to guide researchers in selecting appropriate alignment strategies based on their specific research objectives.

Comparative Performance Analysis

Quantitative Benchmarking of Alignment Modes

Table 1: Performance comparison between one-pass and two-pass alignment modes

Performance Metric	One-Pass Alignment	Two-Pass Alignment	Experimental Context
Uniquely mapped reads	Baseline (reference)	Decreased by 0.4-2%	ENCODE DDX55 KD data [7]
Computational time	Baseline (reference)	Increased by 3-5 minutes per sample	Small test dataset [7]
Significant LSVs detected	Fewer detected	10-15% more detected	GTEx tissue comparison [7]
Reproducibility of unique LSVs	Higher (≈80%)	Lower (≈60-70%)	GTEx 5 vs 5 samples [7]
dPSI correlation between passes	High (r > 0.95)	Moderate (r ≈ 0.85)	Events unique to each method [7]
Novel junction quantification	Lower sensitivity	1.7-fold median read depth improvement	Simulated junction analysis [8]

Table 2: Impact of junction filtering on two-pass alignment performance

Filtration Parameter	Performance Before Filtration	Performance After Filtration	Impact on Results
Low coverage junctions	High junction count	Column 7 < 5 removed	Reduced computational burden [7]
Non-canonical junctions	Included in annotation	Column 5 = 0 removed	Improved specificity [7]
Mitochondrial genes	Included in annotation	chrM junctions removed	Minimal effect on nuclear splicing [7]
Runtime increase	3-5 minutes per sample	1-2 minutes per sample	Improved efficiency [7]
Uniquely mapped reads	1-2% decrease	0.4% decrease	Better mapping statistics [7]

The quantitative comparison reveals a fundamental trade-off between detection sensitivity and measurement reproducibility. Two-pass alignment consistently identifies more splicing events, with one study reporting approximately 1.7-fold deeper median read coverage over novel splice junctions [8]. This enhanced sensitivity stems from the method's ability to align reads with shorter overhangs to junctions discovered during the first pass, effectively reducing the stringency for potentially novel splicing events [8].

However, this increased detection capability comes with significant reproducibility costs. Events uniquely identified through two-pass alignment demonstrate substantially lower reproducibility rates (60-70%) compared to those detected by one-pass methods (approximately 80%) when validated across independent sample sets [7]. This pattern persists across multiple datasets, including GTEx tissue comparisons and ENCODE knockdown experiments, suggesting a fundamental methodological characteristic rather than dataset-specific artifact.

Molecular Implications for Splice Junction Detection

The core technical difference between alignment modes lies in how they handle junction evidence. One-pass alignment requires substantial independent evidence for novel junction support, while two-pass mode incorporates first-pass discoveries as known annotations, effectively reducing the evidence threshold for the same junctions in subsequent alignment. This approach particularly benefits junctions with minimal read overhangs that would otherwise fail alignment thresholds [8].

Potential alignment errors introduced through this process are not random but systematically affect specific transcript characteristics. Short exons, such as the 42-nucleotide exon 6 in Arabidopsis FLM (AT1G77080), demonstrate particularly high rates of misalignment due to insufficient alignment bonuses to overcome intron opening penalties when using standard parameters [15]. Without guidance from reference annotations, as in one-pass mode, only 19.3% of simulated reads aligned correctly to this challenging isoform, while two-pass approaches improved correct alignment to 92.1% by leveraging discovered junctions [15].

Experimental Protocols

Implementation of One-Pass STAR Alignment

Protocol 1: Standard one-pass alignment for reproducible splicing analysis

• Genome Index Preparation: Generate reference indices using well-curated annotations (GENCODE recommended for human data)

• Alignment Execution:

• Quality Assessment:

  • Calculate uniquely mapped read percentage (target >85%)
  • Examine splice junction distribution from SJ.out.tab files
  • Verify expected correlation between technical replicates (>0.95 for expression)

This approach emphasizes measurement stability and is particularly suitable for studies where experimental validation resources are limited or when prioritizing robust differential splicing detection between conditions.

Implementation of Filtered Two-Pass STAR Alignment

Protocol 2: Filtered two-pass alignment for maximal junction discovery

• First Pass Alignment:

• Junction Filtration:

  • Extract high-confidence junctions from SJ.out.tab file
  • Apply filtration thresholds:
    • Minimum junction read support: 5 unique reads (column 7)
    • Remove non-canonical motifs (column 5 ≠ 0)
    • Exclude mitochondrial junctions (column 1 ≠ "chrM")

• Second Pass Alignment:

This filtered approach mitigates the reproducibility challenges of standard two-pass alignment by removing spurious junctions before the second pass, balancing discovery power with analytical reliability [7].

Experimental Validation Framework

Protocol 3: Performance assessment for alignment method selection

• Reproducibility Quantification:

  • Process independent sample replicates through identical pipelines
  • Calculate percentage overlap of significantly changing LSVs (|dPSI| > 0.2, P > 0.95)
  • Target: >75% reproducibility for one-pass, >60% for two-pass unique events

• Sensitivity Benchmarking:

  • Compare detected junction counts against established annotations
  • Quantify known junction recovery rate (expect >90% for highly expressed genes)
  • Document novel junctions per sample as indicator of discovery potential

• Accuracy Assessment:

  • Select representative junctions for experimental validation (RT-PCR)
  • Calculate validation rate (typically >80% for high-confidence junctions)
  • Correlate RNA-seq quantifications with orthogonal methods (Nanostring, qPCR)

This validation framework provides critical empirical data for selecting the optimal alignment strategy based on study-specific requirements and resources.

The Scientist's Toolkit

Table 3: Essential research reagents and computational tools for splicing analysis

Resource	Type/Model	Application Context	Performance Specifications
STAR Aligner	Spliced read mapper	RNA-seq read alignment	Fast processing; splice-aware alignment [16]
MAJIQ	Splicing quantification	Differential splicing analysis	LSV identification and dPSI calculation [7]
GENCODE Annotations	Reference transcriptome	Genome indexing	Comprehensive gene models; regular updates
ENCODE Datasets	Experimental RNA-seq data	Method benchmarking	Well-controlled knockdown studies [7]
GTEx Datasets	Tissue RNA-seq data	Biological variability assessment	Diverse tissue types; multiple donors [7]
2passtools	Junction filtration	Two-pass alignment improvement	Machine-learning-based filtering [15]

The choice between one-pass and two-pass alignment strategies involves fundamental trade-offs between discovery sensitivity and measurement reproducibility. One-pass alignment provides more conservative and reproducible results, making it particularly suitable for hypothesis-driven research focused on robust differential splicing detection. Conversely, two-pass alignment maximizes novel junction discovery, benefiting exploratory studies where comprehensive junction cataloging is prioritized.

Based on empirical evidence, we recommend:

• Use one-pass alignment for clinical applications, biomarker validation studies, and any research requiring high confidence in differential splicing results, as it provides superior reproducibility (≈80% versus 60-70% for two-pass unique events) [7].

• Implement filtered two-pass alignment for exploratory discovery research, novel isoform detection, and studies of poorly annotated transcriptomes, where its enhanced sensitivity (1.7-fold improvement in junction coverage) justifies additional validation requirements [8].

• Apply junction filtration in all two-pass implementations to mitigate reproducibility challenges, using established thresholds for read support and canonical motifs to maintain analytical rigor [7].

• Adopt consistent alignment parameters across studies to ensure comparable results, particularly when integrating datasets from multiple sources or conducting meta-analyses.

As RNA-seq applications continue evolving toward clinical implementation, methodological transparency and reproducibility become increasingly critical. The protocols and benchmarks provided here offer a foundation for robust splicing analysis aligned with rigorous scientific standards required for drug development and clinical research.

The accurate alignment of RNA sequencing (RNA-seq) reads is a foundational step in transcriptomics, enabling the discovery of novel splice junctions and the quantification of gene expression. The Spliced Transcripts Alignment to a Reference (STAR) aligner has been a widely used tool since its development, prized for its high sensitivity in detecting canonical and non-canonical splices. However, the bioinformatics landscape is dynamic, with emerging tools and methodologies posing important questions about relative performance. This Application Note provides a structured comparison of STAR against a selection of modern alignment and quantification tools, focusing on performance metrics critical for novel splice junction detection. We synthesize quantitative evidence from recent studies and provide detailed protocols to guide researchers and drug development professionals in selecting and implementing the optimal workflow for their experimental objectives.

Performance Benchmarking and Quantitative Comparison

STAR vs. Traditional Splice-Aware Aligners

When compared to other alignment-based tools, STAR consistently demonstrates high sensitivity, particularly for splice junction detection. A key differentiator is its performance in identifying novel splice junctions, which is often enhanced by employing a two-pass alignment method [8].

Table 1: Comparison of STAR with other splice-aware aligners

Aligner	Key Strength	Novel Junction Detection	Mapping Rate	Computational Resources	Best Suited For
STAR	High splice junction recall & precision [6] [52]	Excellent, especially with 2-pass mode [8]	High [52]	High memory (~30GB+ for human) [53]	Novel junction discovery, full-length RNA mapping, chimeric transcript detection [6]
HISAT2	Fast runtime, low memory footprint [52]	Good	High, slightly better in some targeted tests [54]	Moderate	Standard differential expression analyses, projects with limited compute resources
TopHat2	—	—	Lower than modern aligners [52]	—	Largely superseded by HISAT2 [52]

STAR vs. Pseudoaligners for Transcript Quantification

Pseudoaligners like Kallisto represent a different algorithmic approach, trading comprehensive genomic mapping for ultra-fast transcript-level quantification.

Table 2: STAR vs. Kallisto feature comparison

Feature	STAR	Kallisto
Core Algorithm	Alignment-based to a reference genome [55]	Pseudoalignment to a transcriptome [55]
Primary Output	Read counts per gene, BAM alignment files [55]	Transcripts per Million (TPM), estimated counts [55]
Key Strength	Discovery of novel splice junctions, fusion genes, and unannotated features [55]	Extremely fast and memory-efficient quantification [55]
Junction Awareness	Directly models and discovers splice junctions from the genome	Relies on a pre-defined transcriptome; cannot discover unannotated junctions
Computational Profile	Slower, high memory usage [55]	Very fast, low memory usage [55]

The choice between these tools depends on the experimental goal. For projects where the objective is the discovery of novel splicing events or fusion transcripts, STAR is the unequivocal choice [55]. However, for large-scale studies focused solely on quantifying expression against a well-annotated transcriptome, Kallisto offers a compelling advantage in speed and resource efficiency [55].

Experimental Protocols for Performance Evaluation

Protocol 1: Two-Pass STAR Alignment for Novel Splice Junction Discovery

The two-pass method increases sensitivity for novel junctions by using information from a first alignment pass to inform the second pass [8].

Procedure:

• First Pass Alignment:
  • Align RNA-seq reads to the reference genome using STAR with standard parameters.
  • Use the -outFilterType BySJout parameter to reduce false positive junctions.
  • Set intron size boundaries with -alignIntronMin 20 and -alignIntronMax 1000000.
  • For a stringent first pass, require a longer overhang for novel junctions, e.g., -alignSJoverhangMin 8.
  • The critical output is the SJ.out.tab file, which contains all detected splice junctions.

• Generate a Novel Junction Database:

  • Combine SJ.out.tab files from all samples in the experiment.
  • Filter the combined list to exclude junctions that are already present in the standard gene annotation (e.g., GENCODE). The remaining junctions are considered "novel" for the purpose of re-indexing.

• Genome Re-indexing:

  • Use the filtered list of novel junctions from Step 2 as an -sjdbFileChrStartEnd input during STAR genome generation.
  • Run the STAR -runMode genomeGenerate command again, including the original GTF annotation and the novel junction file. This creates a splice-aware genome index enriched with sample-specific novel junctions.

• Second Pass Alignment:

  • Re-align all RNA-seq reads to the new genome index generated in Step 3.
  • The aligner will now treat the novel junctions from the first pass as "known," allowing reads to map to them with higher sensitivity and a shorter required overhang (e.g., -alignSJDBoverhangMin 3).
  • The resulting BAM and junction files from this pass are used for downstream quantification and analysis.

Protocol 2: Comparative Workflow for Splicing Tool Performance

This protocol outlines a method to benchmark STAR against other aligners or splicing detection tools using a validated set of known splicing events [54].

Procedure:

• Sample and Data Selection:
  • Obtain RNA-seq data from samples with previously validated splicing events (e.g., exon skipping, alternative splice sites). These can be from public repositories or in-house cell lines or patient samples with orthogonally validated variants.
  • Ensure the dataset includes a range of read lengths (e.g., 50-150bp) and sequencing depths to assess performance across different data qualities.

• Parallel Data Processing:

  • Process the same dataset through multiple alignment workflows.
    • Workflow A: STAR alignment (with and without two-pass mode).
    • Workflow B: HISAT2 alignment.
    • Workflow C: Pseudoalignment with Kallisto (for quantification-only comparison).
  • For alignment-based tools, generate BAM files for downstream analysis.

• Splicing Analysis:

  • Process the BAM files from each aligner through one or more splicing quantification tools, such as rMATS (for exon skipping), MAJIQ (for various event types), or the DEJU workflow [54] [32].
  • Ensure all tools use the same reference genome and annotation.

• Performance Metrics and Validation:

  • Sensitivity (Recall): Calculate the percentage of known, validated splicing events detected by each pipeline (STAR + rMATS vs. HISAT2 + rMATS, etc.).
  • Precision: For novel junction calls, use tools like DeepSplice to classify and filter false positives, or validate a subset via PCR [47].
  • Quantitative Accuracy: Compare the Percent Spliced In (PSI) values calculated by the tools for known events against a validated ground truth [54].

Workflow Visualization

The following diagram illustrates the key decision points and pathways for selecting and applying STAR and emerging tools based on different research goals.

The Scientist's Toolkit

Table 3: Essential research reagents and computational tools for splice junction analysis

Item	Function/Description	Example/Note
STAR Aligner	Spliced alignment of RNA-seq reads to a reference genome.	Version 2.7.10b; requires significant RAM [53].
Reference Genome	Baseline sequence for read alignment.	GRCh38 for human; TAIR10 for A. thaliana [8].
Gene Annotation	Known gene models for guiding alignment and quantifying known features.	GENCODE (human) or Ensembl annotations [8].
SRA Toolkit	Prefetching and converting public RNA-seq data from the NCBI SRA.	Used for fasterq-dump to get FASTQ files [53].
Splicing Quantification Tools	Detecting and quantifying differential splicing events from BAM files.	rMATS (exon skipping), MAJIQ (multiple events) [54].
High-Performance Computing	Computational resources for running resource-intensive aligners.	Cloud (AWS Batch) or HPC clusters; 12+ cores, >32GB RAM recommended [53].

Conclusion

Optimizing STAR alignment parameters, particularly through the implementation of a carefully configured two-pass approach, significantly enhances the detection and quantification of novel splice junctions—a capability with profound implications for understanding disease mechanisms and developing targeted therapies. While this method provides substantial improvements in sensitivity, researchers must balance this with considerations of reproducibility and computational efficiency through appropriate filtering and validation. Future directions should focus on integrating long-read sequencing technologies to fully resolve transcript structures, incorporating deep learning models for improved splice site prediction, and establishing standardized benchmarking frameworks for clinical applications. As splicing-focused therapeutics advance, robust bioinformatic detection of disease-relevant splicing events will become increasingly central to personalized medicine approaches in oncology, neurodegeneration, and genetic disorders.

References

• 1. Five questions on alternative splicing with Omar El Garwany [https://sangerinstitute.blog/2025/08/14/five-questions-on-alternative-splicing-with-omar-el-garwany/]
• 2. Alternative Splicing Regulation in Metabolic Disorders [https://pubmed.ncbi.nlm.nih.gov/40425033/]
• 3. [1306.3466] Sashimi plots: Quantitative visualization of RNA ... [https://ar5iv.labs.arxiv.org/html/1306.3466]
• 4. GLiMMPS: robust statistical model for regulatory variation of alternative... [https://genomebiology.biomedcentral.com/articles/10.1186/gb-2013-14-7-r74]
• 5. Frontiers | RNA - seq science: From raw data to effective... data [https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2023.997383/full]
• 6. STAR: ultrafast universal RNA-seq aligner - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC3530905/]
• 7. How do STAR (settings) affect RNA splicing analysis? [https://biociphers.wordpress.com/2023/11/08/how-do-star-settings-affect-rna-splicing-analysis/]
• 8. Two-pass alignment improves novel splice junction ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5006238/]
• 9. Enhanced detection of splice altering variants in ... [https://www.sciencedirect.com/science/article/abs/pii/S1525157825002387]
• 10. Optimizing RNA-Seq Mapping with STAR [https://www.rna-seqblog.com/optimizing-rna-seq-mapping-with-star/]
• 11. Computational methods for alternative polyadenylation and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12411624/]
• 12. Splice aware aligner - what does it mean? [https://www.biostars.org/p/175454/]
• 13. Alignment with STAR | Introduction to RNA-Seq using high ... [https://hbctraining.github.io/Intro-to-rnaseq-hpc-O2/lessons/03_alignment.html]
• 14. Mapping RNA-seq Reads with STAR - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC4631051/]
• 15. 2passtools: two-pass alignment using machine-learning ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-021-02296-0]
• 16. STAR [https://support.illumina.com/help/BS_App_RNASeq_Alignment_OLH_1000000006112/Content/Source/Informatics/STAR_RNAseq.htm]
• 17. Read alignment | [“Introduction to RNA Sequencing ... [https://huoww07.github.io/Bioinformatics-for-RNA-Seq/lessons/03_Read_Alignment.html]
• 18. Sjdboverhang Option In Star [https://www.biostars.org/p/93883/]
• 19. Confused about sjdbOverhang [https://groups.google.com/g/rna-star/c/h9oh10UlvhI/m/BfSPGivUHmsJ]
• 20. alignSJDBoverhangMin parameter [https://groups.google.com/g/rna-star/c/9mwgYSZeOlo]
• 21. Using STAR for mouse RNA-seq alignment, which ... [https://github.com/alexdobin/STAR/issues/1405]
• 22. Efficient and accurate detection of splice junctions from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6302956/]
• 23. Quantitative analysis of cryptic splicing associated with TDP ... [https://bmcmedgenomics.biomedcentral.com/articles/10.1186/s12920-017-0274-1]
• 24. EI-CoreBioinformatics/portcullis: Splice junction analysis ... [https://github.com/EI-CoreBioinformatics/portcullis]
• 25. Differential Alternative Splicing Analysis with rMATS [https://bioinformatics.ccr.cancer.gov/docs/btep-coding-club/CC2023/rmats/]
• 26. rMATS-turbo – an efficient and flexible computational tool ... [https://www.rna-seqblog.com/rmats-turbo-an-efficient-and-flexible-computational-tool-for-alternative-splicing-analysis-of-large-scale-rna-seq-data/]
• 27. Improved methods for RNAseq-based alternative splicing ... [https://www.nature.com/articles/s41598-021-89938-2]
• 28. Selecting differential splicing methods - PubMed Central - NIH [https://pmc.ncbi.nlm.nih.gov/articles/PMC12308171/]
• 29. Star Aligner Manual: Your Guide to Perfect Alignment [https://san1tar1um.com/star-aligner-manual/]
• 30. Troubleshooting False Positives in Low-to-Medium ... [https://github.com/alexdobin/STAR/issues/1296]
• 31. strand specific control of novel splice junction discovery [https://groups.google.com/g/rna-star/c/knCTreECffI]
• 32. Incorporating exon–exon junction reads enhances differential ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-025-06210-4]
• 33. FineSplice, enhanced splice junction detection and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4005686/]
• 34. Systematic assessment of long-read RNA-seq methods for ... [https://www.nature.com/articles/s41592-024-02298-3]
• 35. Long-read sequencing transcriptome quantification with lr- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11275803/]
• 36. Intron Retention: A Reemerging Paradigm in RNA Biology and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12385487/]
• 37. Cracking rare disorders: a new minimally invasive RNA- ... [https://www.nature.com/articles/s41525-025-00502-7]
• 38. RNA-seq Splicing Profile of the CDH1 Gene and Its Impact ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12562425/]
• 39. Long-read RNA-seq demarcates cis- and trans-directed ... [https://www.nature.com/articles/s41467-025-64605-6]
• 40. MntJULiP and Jutils – differential splicing analysis of RNA- ... [https://www.rna-seqblog.com/mntjulip-and-jutils-differential-splicing-analysis-of-rna-seq-data-with-covariates/]
• 41. ASHG 2025: Torino Identifies mRNA Isoforms in Disease- ... [https://www.genengnews.com/topics/omics/ashg-2025-torino-identifies-mrna-isoforms-in-disease-related-gene-expression/]
• 42. STAR parameters in plants genomes [https://biostar.galaxyproject.org/p/24185/index.html]
• 43. Incorporating exon–exon junction reads enhances ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12288301/]
• 44. Systematically developing a registry of splice-site creating ... [https://www.nature.com/articles/s41467-024-55185-y]
• 45. SpliPath enhances disease gene discovery in case-control ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12570325/]
• 46. A systematic benchmark of Nanopore long-read RNA ... [https://www.nature.com/articles/s41592-025-02623-4]
• 47. Discerning novel splice junctions derived from RNA-seq ... [https://bmcgenomics.biomedcentral.com/articles/10.1186/s12864-018-5350-1]
• 48. Specific splice junction detection in single cells with SICILIAN [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-021-02434-8]
• 49. Machine learning-optimized targeted detection of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11463589/]
• 50. A Beginner's Guide to Analysis of RNA Sequencing Data [https://pmc.ncbi.nlm.nih.gov/articles/PMC6096346/]
• 51. RNA-seq validation: software for selection of reference and ... [https://bmcgenomics.biomedcentral.com/articles/10.1186/s12864-024-10511-y]
• 52. Comparison of Short-Read Sequence Aligners Indicates ... [https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2021.657240/full]
• 53. Performance Analysis and Optimization of the STAR ... [https://arxiv.org/html/2506.12611v1]
• 54. Computational Comparison of Differential Splicing Tools ... [https://www.mdpi.com/1422-0067/26/7/3220]
• 55. Kallisto vs. STAR: Alignment and Quantification of Bulk ... [https://www.elucidata.io/blog/kallisto-vs-star-alignment-and-quantification-of-bulk-rna-seq-data]